Also available: Flash-Next versus Qwen3.8-27B. Full Off, matched xHigh, separate HH2 and provisional XBench profiles retain their measurements and denominators.
Also available: Six Qwen3.8-27B Q6 models in Blender. All 24 retained attempts, the six-model measured summary and separate Codex editorial ratings. Structural /8 scores do not establish visual quality.
Benchmark results explorer
Capture dates refer to the original tests. Video, review, and audit dates are labeled separately. Open a result for software versions, timestamp precision, and recorded settings.
QuantBench / Reviewed: 10 Oct 2026
12/16 GB · QuantBench thinking on
Equal-task scores for seven thinking-capable model files.
Each of 50 tasks has equal weight after averaging its full-cohort and seed 5091/5092 executions. Seed 5090 repeats excluded by aggregation policy. Repeat coverage differs. On/Off samplers and effort controls also differ; this does not isolate a thinking-toggle effect.
7 results
Mean generation tok/s across all completed original-series executions, including seed 5090; excludes prefill · — = not available
12/16 GB · QuantBench thinking on. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Model & configuration
Score / 100
Generation tok/s
Suite minutes
Memory see basis
Evidence
Qwen3.8-27B Heretic GSQ-RCO IQ3_XXSOn; effort differs by model · Q4_0 / Q4_0Campaign captured: Capture, runtime & configuration for Qwen3.8-27B Heretic GSQ-RCO IQ3_XXS
Qwen3.8-27B Heretic GSQ-RCO IQ3_XXS
Test captured
Date precision
Selected request timestamps and task environment snapshots; not an uninterrupted suite duration
llama.cpp extension 2.47.0 · selectedPrismML llama.cpp b10709 · 9a9394a895b96003ca842a6041cb28ac49a108f7llama.cpp · version not recorded
Benchmark
QB3-3.3.8 · QB3-SCORE-3.3.0
Context capacity
65,536 tokens
Thinking
Off; effort differs by model
MTP
Off
K / V cache
Q4_0 / Q4_0
Each of 50 tasks has equal weight after averaging its full-cohort and seed 5091/5092 executions. Seed 5090 repeats excluded by aggregation policy. Repeat coverage differs. On/Off samplers and effort controls also differ; this does not isolate a thinking-toggle effect.
7 results
Mean generation tok/s across all completed original-series executions, including seed 5090; excludes prefill · — = not available
12/16 GB · QuantBench thinking off. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Model & configuration
Score / 100
Generation tok/s
Suite minutes
Memory see basis
Evidence
Qwen3.8-27B Heretic GSQ-RCO IQ3_XXSOff; effort differs by model · Q4_0 / Q4_0Campaign captured: Capture, runtime & configuration for Qwen3.8-27B Heretic GSQ-RCO IQ3_XXS
Qwen3.8-27B Heretic GSQ-RCO IQ3_XXS
Test captured
Date precision
Selected request timestamps and task environment snapshots; not an uninterrupted suite duration
Capture timezone
UTC
Application / version
LM Studio 0.4.25 Build 1
Inference engine / build
llama.cpp · version not recorded
Selected runtime extension
[email protected] Loaded engine build and upstream commit not recorded.
Original temperature 0.7. Adjusted Ministral temperature 0.05 with generic reasoning parsing disabled. This does not isolate a temperature effect. Original Ministral 3B uses a recovered 1 + 49 answer join.
4 results
Mean generation tok/s across this complete 50-task cohort; excludes prefill · — = not available
12/16 GB · QuantBench Instruct original. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Original temperature 0.7. Adjusted Ministral temperature 0.05 with generic reasoning parsing disabled. This does not isolate a temperature effect. Original Ministral 3B uses a recovered 1 + 49 answer join.
3 results
Mean generation tok/s across this complete 50-task cohort; excludes prefill · — = not available
12/16 GB · QuantBench Instruct corrected. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Classic score here is positive-key retrieval percentage, not the 50/50 composite. Reasoning requires ordered retrieval plus a correct result. Formatting can reject readable answers. Captures use repeated fixed prompts; denominators count executions. Historical output cap 4096 and newer uncapped requests are labeled per result. No complete overall is inferred for a capped suite.
11 results
Not aggregated for this profile · — = not available
12/16 GB · Classic 128K retrieval · Q8. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
84 answersstrict parser outcome, not proof of semantic abstention
absent queries
400 questionsclassic128
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
97.60
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
122 answersstrict parser outcome, not proof of semantic abstention
absent queries
400 questionsclassic128
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
97.50
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
60 answersstrict parser outcome, not proof of semantic abstention
absent queries
100 questionsclassic128
Strict parser counts; readable noncompliant answers do not count. Includes one context-capped request; observed coverage is not an eligible complete stage score.
74.20
—
—
—
CappedStrict parser counts; readable noncompliant answers do not count. Includes one context-capped request; observed coverage is not an eligible complete stage score.
Classic score here is positive-key retrieval percentage, not the 50/50 composite. Reasoning requires ordered retrieval plus a correct result. Formatting can reject readable answers. Captures use repeated fixed prompts; denominators count executions. Historical output cap 4096 and newer uncapped requests are labeled per result. No complete overall is inferred for a capped suite.
11 results
Not aggregated for this profile · — = not available
12/16 GB · Classic 240K retrieval · Q8. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
0 answersstrict parser outcome, not proof of semantic abstention
absent queries
80 questionsclassic240
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
51.00
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
20 answersstrict parser outcome, not proof of semantic abstention
absent queries
80 questionsclassic240
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
51.00
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
Gemma 4 12B IT QAT Q4_0Off · Q8_0 / Q8_0Campaign captured: Capture, runtime & configuration for Gemma 4 12B IT QAT Q4_0
Gemma 4 12B IT QAT Q4_0
Test captured
Date precision
Selected request timestamps and task environment snapshots; not an uninterrupted suite duration
Capture timezone
UTC
Application / version
LM Studio 0.4.25 Build 1
Inference engine / build
llama.cpp · version not recorded
Selected runtime extension
[email protected] Loaded engine build and upstream commit not recorded.
Classic score here is positive-key retrieval percentage, not the 50/50 composite. Reasoning requires ordered retrieval plus a correct result. Formatting can reject readable answers. Captures use repeated fixed prompts; denominators count executions. Historical output cap 4096 and newer uncapped requests are labeled per result. No complete overall is inferred for a capped suite.
11 results
Not aggregated for this profile · — = not available
12/16 GB · Reasoning 128K · Q8. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
10.00
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
10.00
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
Classic score here is positive-key retrieval percentage, not the 50/50 composite. Reasoning requires ordered retrieval plus a correct result. Formatting can reject readable answers. Captures use repeated fixed prompts; denominators count executions. Historical output cap 4096 and newer uncapped requests are labeled per result. No complete overall is inferred for a capped suite.
2 results
Not aggregated for this profile · — = not available
12/16 GB · Classic 128K retrieval · Q4. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
120 answersstrict parser outcome, not proof of semantic abstention
absent queries
300 questionsclassic128
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
96.00
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
120 answersstrict parser outcome, not proof of semantic abstention
absent queries
300 questionsclassic128
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
96.00
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
HomHaystack / Reviewed: 10 Oct 2026
12/16 GB · Classic 240K retrieval · Q4
Saved long-context component results. Scores require the benchmark answer format.
Classic score here is positive-key retrieval percentage, not the 50/50 composite. Reasoning requires ordered retrieval plus a correct result. Formatting can reject readable answers. Captures use repeated fixed prompts; denominators count executions. Historical output cap 4096 and newer uncapped requests are labeled per result. No complete overall is inferred for a capped suite.
2 results
Not aggregated for this profile · — = not available
12/16 GB · Classic 240K retrieval · Q4. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
60 answersstrict parser outcome, not proof of semantic abstention
absent queries
60 questionsclassic240
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
54.00
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
0 answersstrict parser outcome, not proof of semantic abstention
absent queries
60 questionsclassic240
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
52.00
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
HomHaystack / Reviewed: 10 Oct 2026
12/16 GB · Reasoning 128K · Q4
Saved long-context component results. Scores require the benchmark answer format.
Classic score here is positive-key retrieval percentage, not the 50/50 composite. Reasoning requires ordered retrieval plus a correct result. Formatting can reject readable answers. Captures use repeated fixed prompts; denominators count executions. Historical output cap 4096 and newer uncapped requests are labeled per result. No complete overall is inferred for a capped suite.
2 results
Not aggregated for this profile · — = not available
12/16 GB · Reasoning 128K · Q4. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
3.33
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
Strict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
1.67
—
—
—
CompleteStrict parser counts; readable noncompliant answers do not count. Repeated fixed prompts and identical saved outputs do not expand distinct-question coverage.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Performance / Reviewed: 10 Oct 2026
12/16 GB · Measured process VRAM · 64K
Highest sampled dedicated model-process GPU memory across available 64K workloads.
llama.cpp extension 2.47.0 · selectedPrismML llama.cpp b10709 · 9a9394a895b96003ca842a6041cb28ac49a108f7llama.cpp · version not recordedllama.cpp extension 2.51.0 · selected
Benchmark
Recorded Windows GPU Process Memory
Context capacity
65,536 tokens
Thinking
Available On/Off workloads; not one paired score
MTP
Off
K / V cache
Q4_0 / Q4_0
RTX 5090 only. Approximately two-second samples may miss transients. Excludes other processes and shared memory. The 2 GiB reserve is a planning choice. 128K values are extrapolations from context probes, not measurements. Configured full offload does not prove actual layer residency.
11 results
Not aggregated for this profile · — = not available
12/16 GB · Measured process VRAM · 64K. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Model & configuration
Score / 100
Generation tok/s
Suite minutes
Memory see basis
Evidence
Qwen3.8-27B Heretic GSQ-RCO IQ3_XXSAvailable On/Off workloads; not one paired score · Q4_0 / Q4_0Campaign captured: Capture, runtime & configuration for Qwen3.8-27B Heretic GSQ-RCO IQ3_XXS
Qwen3.8-27B Heretic GSQ-RCO IQ3_XXS
Test captured
Date precision
Selected request timestamps and task environment snapshots; not an uninterrupted suite duration
Capture timezone
UTC
Application / version
LM Studio 0.4.25 Build 1
Inference engine / build
llama.cpp · version not recorded
Selected runtime extension
[email protected] Loaded engine build and upstream commit not recorded.
OS / environment
win32 10.0.26200
GPU driver version
617.14
CUDA runtime version
Not recorded for this capture
Context capacity
65,536 tokens
K / V cache
Q4_0 / Q4_0
Thinking
Available On/Off workloads; not one paired score
MTP
Off
Memory tool
Windows GPU Process Memory / DedicatedUsage
Sampling interval
Approximately 2 seconds
Fit qualification
Projected from RTX 5090; physical 12/16 GB not tested
Same 50-question suite. ThinkingCap GGUF completed 48/50; its overall is unavailable and its speed is partial. These compare complete configurations, including runtime, cache and template differences. Memory is the highest sampled whole-device reading across selected completions; failed attempts are excluded.
9 results
Arithmetic mean Decode TPS; excludes TTFT. Partial rows are labeled. Effective request TPS and TTFT appear in result details. · — = not available
NVFP4 vs GGUF · QuantBench. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Model & configuration
Score / 100
Generation tok/s
Suite minutes
Memory see basis
Evidence
Base GGUF UD-Q6_KThinking Off · Q8_0 KVCampaign captured: Capture, runtime & configuration for Base GGUF UD-Q6_K · Qwen3.8-27B
Base GGUF UD-Q6_K · Qwen3.8-27B
Test captured
Date precision
Request-wrapper timestamps; exact inference and token start/finish times are not recorded.
Same nine-request reduced workload: 3,405 short / 6,314 long records; weights 30/20/50. OrcaRouter GGUF completed 8/9; no official overall. Keep these scores separate from 262,144-context Legacy results. These compare complete configurations, including runtime, cache and template differences. Memory is the highest sampled whole-device reading across selected completions; failed attempts are excluded.
9 results
Arithmetic mean Decode TPS; excludes TTFT. Partial rows are labeled. Effective request TPS and TTFT appear in result details. · — = not available
NVFP4 vs GGUF · reduced HomHaystack. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Model & configuration
Score / 100
Generation tok/s
Suite minutes
Memory see basis
Evidence
ThinkingCap NVFP4Thinking Off · FP8 E4M3 KVCaptured: Capture, runtime & configuration for ThinkingCap NVFP4 · Qwen3.8-27B
ThinkingCap NVFP4 · Qwen3.8-27B
Test captured
Date precision
First selected request submission to last selected completion; 9/9 completed. Includes gaps between requests; not summed inference duration.
Decode excludes initial wait and rejected MTP proposals. Client committed-token intervals for NVFP4; native per-request generation timing for GGUF.
Suite duration
Not shown; request timing excludes model loads and grading.
MTP acceptance
99.73% · acceptance alone does not establish a speedup
8/9 completed. One request reached the 4,096-token output cap on both allowed attempts. No overall score. Speed and memory cover eight completions only.
Incomplete8/9 completed. One request reached the 4,096-token output cap on both allowed attempts. No overall score. Speed and memory cover eight completions only.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
QuantBench / Reviewed 3 Oct 2026
Uncensored & standard · thinking Off
50 practical tasks across five equally weighted difficulty tiers.
One suite per configuration. Cache precision and thinking effort vary as labeled. VRAM is a sampled device-wide estimate in decimal GB; these are not physical 12/16 GB tests.
29 results
Mean generation tok/s · — = not available
Uncensored & standard · thinking Off. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
One suite per configuration. Cache precision and thinking effort vary as labeled. VRAM is a sampled device-wide estimate in decimal GB; these are not physical 12/16 GB tests.
11 results
Mean generation tok/s · — = not available
Uncensored & standard · thinking checks. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Overall = 30% Classic 128K + 20% Classic 240K + 50% Reasoning 128K. A single 240K run does not establish repeatability. Different QuantBench settings are not carried into this table.
12 results
Mean generation tok/s · — = not available
Uncensored & standard · long context. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
One suite per model. Mean generation includes thinking and final tokens, excluding loading and prefill. Reduced offload is called out in each affected row.
28 results
Mean generation tok/s · — = not available
Swift vs Unsloth · practical tasks. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Inputs: 4,096 / 65,536 / 235,776 tokens. The highest-input stage contributes 60%. Do not compare these totals directly with HS-1.2 or the other context cohort.
5 results
Mean generation tok/s · — = not available
Swift vs Unsloth · 256K. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Inputs: 4,096 / 8,192 / 15,360 tokens. 16 GB fit is projected from RTX 5090 measurements with a 2 GiB reserve. The highest-input stage contributes 60%. Do not compare these totals directly with HS-1.2 or the other context cohort.
5 results
Mean generation tok/s · — = not available
Swift vs Unsloth · 32K. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
None demonstrated both full-suite completion and projected fit within 16 GiB. Partial Unsloth runs have no full-suite score or comparable generation average.
6 results
Mean generation tok/s · — = not available
Swift vs Unsloth · 256K, MTP off. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Model & configuration
Score / 100
Generation tok/s
Suite minutes
Memory see basis
Evidence
Swift IQ2_XSLow · Q4 / Q4Captured: Capture, runtime & configuration for Swift IQ2_XS
Swift IQ2_XS
Test captured
Date precision
Task start and finish include model loading and cleanup; individual inference times may differ.
Capture started
Capture finished
Capture timezone
CDT (UTC−05:00); converted from recorded UTC
Application / version
LM Studio 0.4.24 Build 1
Inference engine / build
llama.cpp · version not recorded
Selected runtime extension
[email protected] Loaded engine build and upstream commit not recorded.
OS / environment
Windows · OS version 10.0.26200
GPU driver version
616.92
CUDA runtime version
Not recorded for this capture
Sampling
Temperature 1 · top-p 0.95 · top-k 20 · seed 5090
FlashAttention / GPU KV
Enabled / Enabled
Parallel requests
1
Evaluation / physical batch
1024 / 256
Min-p / repetition / presence
0 / 1 / 0
GPU offload
Full requested-layer offload recorded
Request watchdog
900 seconds
Input targets
4,096 / 65,536 / 235,776 tokens; three dataset seeds per length
7–9 Sep 2026 · answer datesStage or answer dates in result details
Application / version
LM Studio · version not recorded
Runtime / extension
llama.cpp · version not recorded
Benchmark
QuantBench rubric 2.0.0
Context capacity
262,144 capacity described
Thinking
Not consistently verified
MTP
Not consistently verified
K / V cache
See each row
Historical answer review with incomplete configuration parity. Use the reviewed article values; early workbook grades use different scoring and are not merged here.
7 results
Not reported · — = not available
Original Qwen quant comparison. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Historical absent-query scoring can credit omitted or unparseable answers. Full cache and runtime parity is not established. IQ1_S remains incomplete, not zero.
7 results
Mean generation tok/s · — = not available
Original Qwen · video-era Haystack. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
All seven saved generation-rate observations joined to exact model, ordinal and workbook rate. Submission/final-message times are not exact inference start/end times.
Rate entry 1 submitted
Rate entry 2 submitted
Rate entry 3 submitted
Rate entry 4 submitted
Rate entry 5 submitted
Rate entry 6 submitted
Rate entry 7 submitted
Capture timezone
UTC (saved Unix timestamp conversion)
Application / version
LM Studio · version not recorded
Inference engine / build
llama.cpp · version not recorded
OS / environment
Windows 11 · OS build not recorded
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Cache K / V
Q2/Q3/Q4: Q8_0/Q8_0; Q5: Q4_0/Q4_0
MTP evidence
All four seven-entry speed series record bundled MTP enabled, maximum 1 draft token, and nonzero native draft activity; saved-generation evidence
Saved context capacity
262144
Saved K / V cache
Q8_0 / Q8_0
Saved evaluation / physical batch
2048 / 512
Saved CPU threads
12
Saved load seed
5090
Saved MTP setting
Bundled MTP enabled; maximum 1 draft token
Native draft activity
Nonzero draft and accepted token counts in all seven native generation receipts
Saved FlashAttention / GPU KV
Enabled / enabled
Saved GPU offload ratio
1 requested; actual layer placement not established
Saved parallel requests
1
Saved sampler
Temperature 1; top-k 20; top-p 0.95
Saved min-p field
Entry 1: enabled at 0; entries 2–7: unchecked with stored value 0.05; applied default not established
All seven saved generation-rate observations joined to exact model, ordinal and workbook rate. Submission/final-message times are not exact inference start/end times.
Rate entry 1 submitted
Rate entry 2 submitted
Rate entry 3 submitted
Rate entry 4 submitted
Rate entry 5 submitted
Rate entry 6 submitted
Rate entry 7 submitted
Capture timezone
UTC (saved Unix timestamp conversion)
Application / version
LM Studio · version not recorded
Inference engine / build
llama.cpp · version not recorded
OS / environment
Windows 11 · OS build not recorded
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Cache K / V
Q2/Q3/Q4: Q8_0/Q8_0; Q5: Q4_0/Q4_0
MTP evidence
All four seven-entry speed series record bundled MTP enabled, maximum 1 draft token, and nonzero native draft activity; saved-generation evidence
Saved context capacity
262144
Saved K / V cache
Q8_0 / Q8_0
Saved evaluation / physical batch
2048 / 512
Saved CPU threads
12
Saved load seed
5090
Saved MTP setting
Bundled MTP enabled; maximum 1 draft token
Native draft activity
Nonzero draft and accepted token counts in all seven native generation receipts
Saved FlashAttention / GPU KV
Enabled / enabled
Saved GPU offload ratio
1 requested; actual layer placement not established
Saved parallel requests
1
Saved sampler
Temperature 1; top-k 20; top-p 0.95
Saved min-p field
Entry 1: enabled at 0; entries 2–7: unchecked with stored value 0.05; applied default not established
All seven saved generation-rate observations joined to exact model, ordinal and workbook rate. Submission/final-message times are not exact inference start/end times.
Rate entry 1 submitted
Rate entry 2 submitted
Rate entry 3 submitted
Rate entry 4 submitted
Rate entry 5 submitted
Rate entry 6 submitted
Rate entry 7 submitted
Capture timezone
UTC (saved Unix timestamp conversion)
Application / version
LM Studio · version not recorded
Inference engine / build
llama.cpp · version not recorded
OS / environment
Windows 11 · OS build not recorded
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Cache K / V
Q2/Q3/Q4: Q8_0/Q8_0; Q5: Q4_0/Q4_0
MTP evidence
All four seven-entry speed series record bundled MTP enabled, maximum 1 draft token, and nonzero native draft activity; saved-generation evidence
Saved context capacity
262144
Saved K / V cache
Q8_0 / Q8_0
Saved evaluation / physical batch
2048 / 512
Saved CPU threads
12
Saved load seed
5090
Saved MTP setting
Bundled MTP enabled; maximum 1 draft token
Native draft activity
Nonzero draft and accepted token counts in all seven native generation receipts
Saved FlashAttention / GPU KV
Enabled / enabled
Saved GPU offload ratio
1 requested; actual layer placement not established
Saved parallel requests
1
Saved sampler
Temperature 1; top-k 20; top-p 0.95
Saved min-p field
Entry 1: enabled at 0; entries 2–7: unchecked with stored value 0.05; applied default not established
All seven saved generation-rate observations joined to exact model, ordinal and workbook rate. Submission/final-message times are not exact inference start/end times.
Rate entry 1 submitted
Rate entry 2 submitted
Rate entry 3 submitted
Rate entry 4 submitted
Rate entry 5 submitted
Rate entry 6 submitted
Rate entry 7 submitted
Capture timezone
UTC (saved Unix timestamp conversion)
Application / version
LM Studio · version not recorded
Inference engine / build
llama.cpp · version not recorded
OS / environment
Windows 11 · OS build not recorded
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Cache K / V
Q2/Q3/Q4: Q8_0/Q8_0; Q5: Q4_0/Q4_0
MTP evidence
All four seven-entry speed series record bundled MTP enabled, maximum 1 draft token, and nonzero native draft activity; saved-generation evidence
Saved context capacity
262144
Saved K / V cache
Q4_0 / Q4_0
Saved evaluation / physical batch
2048 / 512
Saved CPU threads
12
Saved load seed
5090
Saved MTP setting
Bundled MTP enabled; maximum 1 draft token
Native draft activity
Nonzero draft and accepted token counts in all seven native generation receipts
Saved FlashAttention / GPU KV
Enabled / enabled
Saved GPU offload ratio
1 requested; actual layer placement not established
Saved parallel requests
1
Saved sampler
Temperature 1; top-k 20; top-p 0.95
Saved min-p field
All seven entries: unchecked, stored value 0.05; applied default not established
A separate campaign and benchmark version. Do not merge into the MTP3 or thinking-enabled rankings. Device estimates are not exclusive model allocation.
1 results
Median generation tok/s · — = not available
byteshape IQ3_XS · practical tasks. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
A separate campaign and benchmark version. Do not merge into the MTP3 or thinking-enabled rankings. Device estimates are not exclusive model allocation.
1 results
Median generation tok/s · — = not available
byteshape IQ3_XS · long context. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
vLLM 0.27.1LM Studio · version not recordedvLLM · version not recorded
Runtime / extension
vLLM 0.27.1llama.cpp · version not recordedvLLM · version not recorded
Benchmark
Timing audit baseline
Context capacity
Frozen campaign profile
Thinking
See per-result settings
MTP
See per-result settings
K / V cache
Backend-specific
The audit confirmed a real configuration-specific slowdown, not a prefill/decode denominator error. Later repair captures are separate from this historical series. Exact model/cache parity is not established here; this is not a universal format verdict.
10 results
Mean native decode tok/s · — = not available
NVFP4 vs GGUF · original speed baseline. Scores are out of 100; Edition-specific benchmark score. Time basis: Suite minutes.
Model & configuration
Score / 100
Generation tok/s
Suite minutes
Memory see basis
Evidence
Base Qwen NVFP4 · QuantBenchThinking Off · Backend-specificCampaign captured: Capture, runtime & configuration for Base Qwen NVFP4 · QuantBench
Base Qwen NVFP4 · QuantBench
Test captured
Date precision
Request-wrapper timestamps; exact inference and token start/finish times are not recorded.
Historical native decode measurement. Later repair captures are separate.
—
81.62
—
—
Audit baselineHistorical native decode measurement. Later repair captures are separate.
ThinkingCap NVFP4 · QuantBenchThinking not recorded · Backend-specificCaptured: Not recordedCapture, runtime & configuration for ThinkingCap NVFP4 · QuantBench
ThinkingCap NVFP4 · QuantBench
Test captured
Not recorded
Date precision
No valid benchmark result; capture date not recorded.
Unavailable in the original campaign. The later repaired NVFP4 campaign has separate completed results.
—
—
—
—
UnavailableUnavailable in the original campaign. The later repaired NVFP4 campaign has separate completed results.
ThinkingCap NVFP4 · HomHaystackThinking not recorded · Backend-specificCaptured: Not recordedCapture, runtime & configuration for ThinkingCap NVFP4 · HomHaystack
ThinkingCap NVFP4 · HomHaystack
Test captured
Not recorded
Date precision
No valid benchmark result; capture date not recorded.
Huihui mean score uses three measured scenes; its fourth score is unavailable. Mean time uses all four retained attempts, including its 45-minute crash deadline. Token counts sum repeated input context and generated tokens.
6 results
Generation rate not included; elapsed time includes tool and render work · — = not available
Blender Q6 · four-attempt measured summary. Scores are out of 8; Mean automated structural checks; measured-scene denominator in each record. Time basis: Mean attempt minutes.
Own-repository projector; screenshot transport verified, image understanding not established
Scope
One mug-building task; does not rank general coding, reasoning or other Blender tasks.
Scenes measured
3/4
Saved model images
4/4
Mean total tokens
472,554.50
Tool errors / calls
5 / 72
Individual scene scores
8, 7, 8, Unavailable
All measurements and denominators (7)
retained attempts
4 attemptsRetained comparison
scored scenes
3 scenesComplete eight-check scores
saved model images
4 imagesValid model-made PNGs
mean total tokens
472,554.5 tokensInput plus generated tokens, including repeated context
tool calls
72 callsSum over four retained attempts
tool errors
5 errorsSum over four retained attempts
mean attempt seconds
967.374 secondsFour attempts; excludes loading, preflight and post-run checks/captures
Score mean covers 3/4 scenes. Run 4 score unavailable after a model-caused crash; its PNG survived. Mean time includes the recorded 45-minute deadline and waiting after the crash. One Windows restart redo is excluded.
7.67
—
16.12
—
Partial · scoredScore mean covers 3/4 scenes. Run 4 score unavailable after a model-caused crash; its PNG survived. Mean time includes the recorded 45-minute deadline and waiting after the crash. One Windows restart redo is excluded.
23 scenes have complete eight-check scores; all 24 attempts saved a model PNG. Huihui run 4 lost its final scene after saving the image. Its missing score is null. The infrastructure-interrupted attempt is excluded.
24 results
Generation rate not included; elapsed time includes tool and render work · — = not available
Blender Q6 · all 24 retained attempts. Scores are out of 8; Automated structural checks per retained attempt; not visual quality. Time basis: Attempt minutes.
Model & configuration
Score / 8
Generation tok/s
Attempt minutes
Memory see basis
Evidence
ThinkingCap · run 1On · native template · Q8_0 / Q8_0 requestedCaptured: Capture, runtime & configuration for ThinkingCap Qwen3.8-27B Q6_K · run 1
ThinkingCap Qwen3.8-27B Q6_K · run 1
Test captured
Date precision
Exact retained model-attempt interval; excludes post-run scene check/capture
Capture started
Capture finished
Capture timezone
UTC
Application / version
LM Studio 0.4.25 Build 1
Inference engine / build
Not recorded
OS / environment
Windows; NVIDIA GeForce RTX 5090 confirmed in the joined model-load readback; exact OS build and GPU driver not established
Own-repository projector; screenshot transport verified, image understanding not established
Scope
One mug-building task; does not rank general coding, reasoning or other Blender tasks.
Total tokens
510,316
Tool errors / calls
2 / 20
Saved model image
1/1
Post-run scene capture
Unavailable after crash
All measurements and denominators (6)
total tokens
510,316 tokensThis retained attempt; total tokens include repeated input context
completion tokens
41,005 tokensThis retained attempt; total tokens include repeated input context
tool calls
20 callsThis retained attempt; total tokens include repeated input context
tool errors
2 errorsThis retained attempt; total tokens include repeated input context
saved model images
1 imagesValid model-made PNG
scored scenes
0 scenesComplete eight-check score available
Final scene score unavailable after a model-caused Blender crash. Saved PNG survived. Time includes the 45-minute deadline and post-crash waiting.
—
—
45.00
—
UnavailableFinal scene score unavailable after a model-caused Blender crash. Saved PNG survived. Time includes the 45-minute deadline and post-crash waiting.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Blender tool calling / Captured 7 October 2026 UTC
Blender Q6 · Codex editorial ratings
Historical subjective judgments, separate from measured structural scores. The later visual audit identified additional handle defects.
Post-hoc Codex ratings, not objective measurements or a blinded panel. Historical values are retained with attribution. The 7.6 vs 7.4 display gap does not establish a decisive winner; changing weights changes the order.
6 results
Generation rate not included; elapsed time includes tool and render work · — = not available
Blender Q6 · Codex editorial ratings. Scores are out of 10; Subjective Codex editorial rating; unrounded weighted value. Time basis: Not applicable.
Own-repository projector; screenshot transport verified, image understanding not established
Scope
One mug-building task; does not rank general coding, reasoning or other Blender tasks.
Visual / workflow / efficiency
7.25 / 9 / 7 (each /10)
Weights
70% visual · 20% workflow · 10% efficiency
Article display
7.6 /10; rounded to one decimal
All measurements and denominators (3)
visual mean
7.25 points / 10Subjective visual average; all four images equally weighted
workflow rating
9 points / 10Subjective workflow assessment
efficiency rating
7 points / 10Comparative judgment informed by time and token use
Historical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
7.58
—
—
—
Subjective editorialHistorical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
Own-repository projector; screenshot transport verified, image understanding not established
Scope
One mug-building task; does not rank general coding, reasoning or other Blender tasks.
Visual / workflow / efficiency
6.50 / 9 / 10 (each /10)
Weights
70% visual · 20% workflow · 10% efficiency
Article display
7.4 /10; rounded to one decimal
All measurements and denominators (3)
visual mean
6.5 points / 10Subjective visual average; all four images equally weighted
workflow rating
9 points / 10Subjective workflow assessment
efficiency rating
10 points / 10Comparative judgment informed by time and token use
Historical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
7.35
—
—
—
Subjective editorialHistorical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
Own-repository projector; screenshot transport verified, image understanding not established
Scope
One mug-building task; does not rank general coding, reasoning or other Blender tasks.
Visual / workflow / efficiency
5.75 / 9 / 9 (each /10)
Weights
70% visual · 20% workflow · 10% efficiency
Article display
6.7 /10; rounded to one decimal
All measurements and denominators (3)
visual mean
5.75 points / 10Subjective visual average; all four images equally weighted
workflow rating
9 points / 10Subjective workflow assessment
efficiency rating
9 points / 10Comparative judgment informed by time and token use
Historical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
6.72
—
—
—
Subjective editorialHistorical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
Own-repository projector; screenshot transport verified, image understanding not established
Scope
One mug-building task; does not rank general coding, reasoning or other Blender tasks.
Visual / workflow / efficiency
5.75 / 9 / 2 (each /10)
Weights
70% visual · 20% workflow · 10% efficiency
Article display
6.0 /10; rounded to one decimal
All measurements and denominators (3)
visual mean
5.75 points / 10Subjective visual average; all four images equally weighted
workflow rating
9 points / 10Subjective workflow assessment
efficiency rating
2 points / 10Comparative judgment informed by time and token use
Historical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
6.03
—
—
—
Subjective editorialHistorical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
Own-repository projector; screenshot transport verified, image understanding not established
Scope
One mug-building task; does not rank general coding, reasoning or other Blender tasks.
Visual / workflow / efficiency
6.50 / 4 / 5 (each /10)
Weights
70% visual · 20% workflow · 10% efficiency
Article display
5.9 /10; rounded to one decimal
All measurements and denominators (3)
visual mean
6.5 points / 10Subjective visual average; all four images equally weighted
workflow rating
4 points / 10Subjective workflow assessment
efficiency rating
5 points / 10Comparative judgment informed by time and token use
Historical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
5.85
—
—
—
Subjective editorialHistorical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
Own-repository projector; screenshot transport verified, image understanding not established
Scope
One mug-building task; does not rank general coding, reasoning or other Blender tasks.
Visual / workflow / efficiency
2.25 / 5 / 7 (each /10)
Weights
70% visual · 20% workflow · 10% efficiency
Article display
3.3 /10; rounded to one decimal
All measurements and denominators (3)
visual mean
2.25 points / 10Subjective visual average; all four images equally weighted
workflow rating
5 points / 10Subjective workflow assessment
efficiency rating
7 points / 10Comparative judgment informed by time and token use
Historical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
3.27
—
—
—
Subjective editorialHistorical post-hoc Codex judgment. See the article for later handle-defect corrections; no numerical ratings were silently revised.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · QuantBench
mean of 50 normalized per-question rubric values, points/100
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · QuantBench. Scores are out of 100; mean of 50 normalized per-question rubric values, points/100. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
2026-10-10T20:21:17.2996256Z to 2026-10-10T20:29:30.0420554Z
Wall measurement methods
{"timestampDifference": 50}
Expert file I/O scope
Zero decode expert reads does not imply zero whole-request storage activity. Engine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
All measurements and denominators (41)
requests
50 responsesOriginal complete Off suite
normalized score
64.433463 points_per_100Full Off suite. mean of per-question normalized rubric points
normalized score denominator
50 countFull Off suite
normalized score by tier extreme
63 points_per_100Full Off suite
normalized score by tier extreme denominator
10 countFull Off suite
normalized score by tier extreme plus
75.522024 points_per_100Full Off suite
normalized score by tier extreme plus denominator
10 countFull Off suite
normalized score by tier extreme plus plus
49.95328 points_per_100Full Off suite
normalized score by tier extreme plus plus denominator
10 countFull Off suite
normalized score by tier extreme plus plus plus
55.19201 points_per_100Full Off suite
normalized score by tier extreme plus plus plus denominator
10 countFull Off suite
normalized score by tier original
78.5 points_per_100Full Off suite
normalized score by tier original denominator
10 countFull Off suite
capped
0 countOriginal complete Off suite; capped and empty may overlap
empty final
0 countOriginal complete Off suite; capped and empty may overlap
overlap capped and empty
0 countOriginal complete Off suite; capped and empty may overlap
50 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
145 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.434324 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.038609 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
66.55402 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
107.516575 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.183594 GiBsampled whole host free memory; not model allocation
native decode file blobs sum
0 file blobsNative decode counters only
native decode file mb
0 MBNative decode counters only
whole run cumulative expert file blobs last
9,700 file blobsEngine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
64.43
137.39
7.86
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
2026-10-10T21:00:02.5679491Z to 2026-10-10T21:23:15.6729196Z
Wall measurement methods
{"timestampDifference": 50}
Expert file I/O scope
Zero decode expert reads does not imply zero whole-request storage activity. Engine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
All measurements and denominators (41)
requests
50 responsesOriginal complete Off suite
normalized score
61.552429 points_per_100Full Off suite. mean of per-question normalized rubric points
normalized score denominator
50 countFull Off suite
normalized score by tier extreme
58 points_per_100Full Off suite
normalized score by tier extreme denominator
10 countFull Off suite
normalized score by tier extreme plus
64.992857 points_per_100Full Off suite
normalized score by tier extreme plus denominator
10 countFull Off suite
normalized score by tier extreme plus plus
50.906916 points_per_100Full Off suite
normalized score by tier extreme plus plus denominator
10 countFull Off suite
normalized score by tier extreme plus plus plus
60.362373 points_per_100Full Off suite
normalized score by tier extreme plus plus plus denominator
10 countFull Off suite
normalized score by tier original
73.5 points_per_100Full Off suite
normalized score by tier original denominator
10 countFull Off suite
capped
0 countOriginal complete Off suite; capped and empty may overlap
empty final
0 countOriginal complete Off suite; capped and empty may overlap
overlap capped and empty
0 countOriginal complete Off suite; capped and empty may overlap
50 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
408 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.441222 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
4.943163 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
46.799942 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
78.607334 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.608398 GiBsampled whole host free memory; not model allocation
native decode file blobs sum
0 file blobsNative decode counters only
native decode file mb
0 MBNative decode counters only
whole run cumulative expert file blobs last
14,702 file blobsEngine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
61.55
54.50
22.75
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
2026-10-10T20:10:45.7177705Z to 2026-10-10T20:17:03.7444947Z
Wall measurement methods
{"timestampDifference": 50}
Expert file I/O scope
Zero decode expert reads does not imply zero whole-request storage activity. Engine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
All measurements and denominators (41)
requests
50 responsesOriginal complete Off suite
normalized score
60.246708 points_per_100Full Off suite. mean of per-question normalized rubric points
normalized score denominator
50 countFull Off suite
normalized score by tier extreme
43 points_per_100Full Off suite
normalized score by tier extreme denominator
10 countFull Off suite
normalized score by tier extreme plus
72.843155 points_per_100Full Off suite
normalized score by tier extreme plus denominator
10 countFull Off suite
normalized score by tier extreme plus plus
46.538199 points_per_100Full Off suite
normalized score by tier extreme plus plus denominator
10 countFull Off suite
normalized score by tier extreme plus plus plus
55.852185 points_per_100Full Off suite
normalized score by tier extreme plus plus plus denominator
10 countFull Off suite
normalized score by tier original
83 points_per_100Full Off suite
normalized score by tier original denominator
10 countFull Off suite
capped
0 countOriginal complete Off suite; capped and empty may overlap
empty final
0 countOriginal complete Off suite; capped and empty may overlap
overlap capped and empty
0 countOriginal complete Off suite; capped and empty may overlap
50 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
113 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.379837 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
4.621653 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
72.459652 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
113.275581 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.143555 GiBsampled whole host free memory; not model allocation
native decode file blobs sum
0 file blobsNative decode counters only
native decode file mb
0 MBNative decode counters only
whole run cumulative expert file blobs last
10,150 file blobsEngine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
60.25
178.64
5.79
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
2026-10-10T20:35:17.4134901Z to 2026-10-10T20:46:12.9996066Z
Wall measurement methods
{"timestampDifference": 50}
Expert file I/O scope
Zero decode expert reads does not imply zero whole-request storage activity. Engine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
All measurements and denominators (41)
requests
50 responsesOriginal complete Off suite
normalized score
58.313312 points_per_100Full Off suite. mean of per-question normalized rubric points
normalized score denominator
50 countFull Off suite
normalized score by tier extreme
60.5 points_per_100Full Off suite
normalized score by tier extreme denominator
10 countFull Off suite
normalized score by tier extreme plus
54.541468 points_per_100Full Off suite
normalized score by tier extreme plus denominator
10 countFull Off suite
normalized score by tier extreme plus plus
43.280157 points_per_100Full Off suite
normalized score by tier extreme plus plus denominator
10 countFull Off suite
normalized score by tier extreme plus plus plus
59.744936 points_per_100Full Off suite
normalized score by tier extreme plus plus plus denominator
10 countFull Off suite
normalized score by tier original
73.5 points_per_100Full Off suite
normalized score by tier original denominator
10 countFull Off suite
capped
0 countOriginal complete Off suite; capped and empty may overlap
empty final
0 countOriginal complete Off suite; capped and empty may overlap
overlap capped and empty
0 countOriginal complete Off suite; capped and empty may overlap
50 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
193 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.439387 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.382659 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
41.408524 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
99.648323 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.285156 GiBsampled whole host free memory; not model allocation
native decode file blobs sum
0 file blobsNative decode counters only
native decode file mb
0 MBNative decode counters only
whole run cumulative expert file blobs last
8,950 file blobsEngine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
58.31
111.14
10.30
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
50 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
576 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.061998 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.138656 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.176498 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
121.293491 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.170898 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
57.61
63.78
19.69
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
2026-10-10T20:01:03.1865267Z to 2026-10-10T20:05:51.4053326Z
Wall measurement methods
{"timestampDifference": 50}
Expert file I/O scope
Zero decode expert reads does not imply zero whole-request storage activity. Engine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
All measurements and denominators (41)
requests
50 responsesOriginal complete Off suite
normalized score
55.514834 points_per_100Full Off suite. mean of per-question normalized rubric points
normalized score denominator
50 countFull Off suite
normalized score by tier extreme
47 points_per_100Full Off suite
normalized score by tier extreme denominator
10 countFull Off suite
normalized score by tier extreme plus
52.200099 points_per_100Full Off suite
normalized score by tier extreme plus denominator
10 countFull Off suite
normalized score by tier extreme plus plus
44.384063 points_per_100Full Off suite
normalized score by tier extreme plus plus denominator
10 countFull Off suite
normalized score by tier extreme plus plus plus
51.490009 points_per_100Full Off suite
normalized score by tier extreme plus plus plus denominator
10 countFull Off suite
normalized score by tier original
82.5 points_per_100Full Off suite
normalized score by tier original denominator
10 countFull Off suite
capped
0 countOriginal complete Off suite; capped and empty may overlap
empty final
0 countOriginal complete Off suite; capped and empty may overlap
overlap capped and empty
0 countOriginal complete Off suite; capped and empty may overlap
50 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
87 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.302977 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.176377 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.799217 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
129.970631 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.101563 GiBsampled whole host free memory; not model allocation
native decode file blobs sum
0 file blobsNative decode counters only
native decode file mb
0 MBNative decode counters only
whole run cumulative expert file blobs last
9,650 file blobsEngine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
55.51
194.10
4.59
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
50 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
883 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.061993 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.155458 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
86.840939 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
124.876369 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.604492 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
51.62
76.90
30.28
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · Legacy96K
tier means weighted Classic short .3, Classic long .2, Reasoning short .5
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · Legacy96K. Scores are out of 100; tier means weighted Classic short .3, Classic long .2, Reasoning short .5. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
2026-10-10T22:47:15.0412862Z to 2026-10-10T22:55:35.1367720Z
Wall measurement methods
{"recordedMonotonic": 9}
Expert file I/O scope
Zero decode expert reads does not imply zero whole-request storage activity. Engine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
All measurements and denominators (46)
requests
9 responsesOriginal complete Off suite
score
100 points_per_100Full Off suite
tier scores classic short
100 points_per_100Full Off suite
tier scores classic short denominator
5 countFull Off suite
tier scores classic short weight
0.3 weightFull Off suite
tier scores classic long 96 k
100 points_per_100Full Off suite
tier scores classic long 96 k denominator
1 countFull Off suite
tier scores classic long 96 k weight
0.2 weightFull Off suite
tier scores reasoning short
100 points_per_100Full Off suite
tier scores reasoning short denominator
3 countFull Off suite
tier scores reasoning short weight
0.5 weightFull Off suite
retrieval records
780 countFull Off suite
input tokens by tier classic short minimum
59,236 tokensFull Off suite
input tokens by tier classic short maximum
59,343 tokensFull Off suite
input tokens by tier classic long 96 k minimum
91,450 tokensFull Off suite
input tokens by tier classic long 96 k maximum
91,450 tokensFull Off suite
input tokens by tier reasoning short minimum
58,582 tokensFull Off suite
input tokens by tier reasoning short maximum
58,603 tokensFull Off suite
capped
0 countOriginal complete Off suite; capped and empty may overlap
empty final
0 countOriginal complete Off suite; capped and empty may overlap
overlap capped and empty
0 countOriginal complete Off suite; capped and empty may overlap
9 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
148 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.409485 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.594551 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
44.797695 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
76.830532 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.625 GiBsampled whole host free memory; not model allocation
native decode file blobs sum
0 file blobsNative decode counters only
native decode file mb
0 MBNative decode counters only
whole run cumulative expert file blobs last
10,665 file blobsEngine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
100.00
71.87
7.84
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
9 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
199 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.061893 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.138489 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.299988 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
123.640617 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
11.198242 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
100.00
62.51
6.75
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
9 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
222 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062315 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.107533 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.761974 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
120.275406 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
6.766602 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
99.14
53.16
7.54
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
2026-10-10T22:07:38.9622672Z to 2026-10-10T22:20:12.4593625Z
Wall measurement methods
{"recordedMonotonic": 9}
Expert file I/O scope
Zero decode expert reads does not imply zero whole-request storage activity. Engine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
All measurements and denominators (46)
requests
9 responsesOriginal complete Off suite
score
99.016667 points_per_100Full Off suite
tier scores classic short
99.5 points_per_100Full Off suite
tier scores classic short denominator
5 countFull Off suite
tier scores classic short weight
0.3 weightFull Off suite
tier scores classic long 96 k
100 points_per_100Full Off suite
tier scores classic long 96 k denominator
1 countFull Off suite
tier scores classic long 96 k weight
0.2 weightFull Off suite
tier scores reasoning short
98.333333 points_per_100Full Off suite
tier scores reasoning short denominator
3 countFull Off suite
tier scores reasoning short weight
0.5 weightFull Off suite
retrieval records
780 countFull Off suite
input tokens by tier classic short minimum
59,236 tokensFull Off suite
input tokens by tier classic short maximum
59,343 tokensFull Off suite
input tokens by tier classic long 96 k minimum
91,450 tokensFull Off suite
input tokens by tier classic long 96 k maximum
91,450 tokensFull Off suite
input tokens by tier reasoning short minimum
58,582 tokensFull Off suite
input tokens by tier reasoning short maximum
58,603 tokensFull Off suite
capped
0 countOriginal complete Off suite; capped and empty may overlap
empty final
0 countOriginal complete Off suite; capped and empty may overlap
overlap capped and empty
0 countOriginal complete Off suite; capped and empty may overlap
9 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
229 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.273838 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
4.994389 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
77.687557 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
108.677658 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.178711 GiBsampled whole host free memory; not model allocation
native decode file blobs sum
0 file blobsNative decode counters only
native decode file mb
0 MBNative decode counters only
whole run cumulative expert file blobs last
2,106 file blobsEngine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
99.02
202.70
12.13
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
2026-10-10T22:20:36.0999609Z to 2026-10-10T22:32:05.3950551Z
Wall measurement methods
{"recordedMonotonic": 9}
Expert file I/O scope
Zero decode expert reads does not imply zero whole-request storage activity. Engine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
All measurements and denominators (46)
requests
9 responsesOriginal complete Off suite
score
97.606667 points_per_100Full Off suite
tier scores classic short
94.8 points_per_100Full Off suite
tier scores classic short denominator
5 countFull Off suite
tier scores classic short weight
0.3 weightFull Off suite
tier scores classic long 96 k
100 points_per_100Full Off suite
tier scores classic long 96 k denominator
1 countFull Off suite
tier scores classic long 96 k weight
0.2 weightFull Off suite
tier scores reasoning short
98.333333 points_per_100Full Off suite
tier scores reasoning short denominator
3 countFull Off suite
tier scores reasoning short weight
0.5 weightFull Off suite
retrieval records
780 countFull Off suite
input tokens by tier classic short minimum
59,236 tokensFull Off suite
input tokens by tier classic short maximum
59,343 tokensFull Off suite
input tokens by tier classic long 96 k minimum
91,450 tokensFull Off suite
input tokens by tier classic long 96 k maximum
91,450 tokensFull Off suite
input tokens by tier reasoning short minimum
58,582 tokensFull Off suite
input tokens by tier reasoning short maximum
58,603 tokensFull Off suite
capped
1 countOriginal complete Off suite; capped and empty may overlap
empty final
0 countOriginal complete Off suite; capped and empty may overlap
overlap capped and empty
0 countOriginal complete Off suite; capped and empty may overlap
9 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
210 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.256529 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
4.868991 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
70.572144 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
101.287609 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.412109 GiBsampled whole host free memory; not model allocation
native decode file blobs sum
0 file blobsNative decode counters only
native decode file mb
0 MBNative decode counters only
whole run cumulative expert file blobs last
1,746 file blobsEngine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
97.61
164.48
11.02
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
2026-10-10T21:59:16.0932395Z to 2026-10-10T22:07:17.6011572Z
Wall measurement methods
{"recordedMonotonic": 9}
Expert file I/O scope
Zero decode expert reads does not imply zero whole-request storage activity. Engine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
All measurements and denominators (46)
requests
9 responsesOriginal complete Off suite
score
97.493333 points_per_100Full Off suite
tier scores classic short
97.2 points_per_100Full Off suite
tier scores classic short denominator
5 countFull Off suite
tier scores classic short weight
0.3 weightFull Off suite
tier scores classic long 96 k
100 points_per_100Full Off suite
tier scores classic long 96 k denominator
1 countFull Off suite
tier scores classic long 96 k weight
0.2 weightFull Off suite
tier scores reasoning short
96.666667 points_per_100Full Off suite
tier scores reasoning short denominator
3 countFull Off suite
tier scores reasoning short weight
0.5 weightFull Off suite
retrieval records
780 countFull Off suite
input tokens by tier classic short minimum
59,236 tokensFull Off suite
input tokens by tier classic short maximum
59,343 tokensFull Off suite
input tokens by tier classic long 96 k minimum
91,450 tokensFull Off suite
input tokens by tier classic long 96 k maximum
91,450 tokensFull Off suite
input tokens by tier reasoning short minimum
58,582 tokensFull Off suite
input tokens by tier reasoning short maximum
58,603 tokensFull Off suite
capped
1 countOriginal complete Off suite; capped and empty may overlap
empty final
0 countOriginal complete Off suite; capped and empty may overlap
overlap capped and empty
0 countOriginal complete Off suite; capped and empty may overlap
9 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
150 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.210491 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.599798 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
94.808849 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
126.098759 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.335938 GiBsampled whole host free memory; not model allocation
native decode file blobs sum
0 file blobsNative decode counters only
native decode file mb
0 MBNative decode counters only
whole run cumulative expert file blobs last
1,728 file blobsEngine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
97.49
207.72
7.74
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
2026-10-10T22:32:24.3893215Z to 2026-10-10T22:46:56.9432175Z
Wall measurement methods
{"recordedMonotonic": 9}
Expert file I/O scope
Zero decode expert reads does not imply zero whole-request storage activity. Engine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
All measurements and denominators (46)
requests
9 responsesOriginal complete Off suite
score
82.833333 points_per_100Full Off suite
tier scores classic short
100 points_per_100Full Off suite
tier scores classic short denominator
5 countFull Off suite
tier scores classic short weight
0.3 weightFull Off suite
tier scores classic long 96 k
97.5 points_per_100Full Off suite
tier scores classic long 96 k denominator
1 countFull Off suite
tier scores classic long 96 k weight
0.2 weightFull Off suite
tier scores reasoning short
66.666667 points_per_100Full Off suite
tier scores reasoning short denominator
3 countFull Off suite
tier scores reasoning short weight
0.5 weightFull Off suite
retrieval records
780 countFull Off suite
input tokens by tier classic short minimum
59,236 tokensFull Off suite
input tokens by tier classic short maximum
59,343 tokensFull Off suite
input tokens by tier classic long 96 k minimum
91,450 tokensFull Off suite
input tokens by tier classic long 96 k maximum
91,450 tokensFull Off suite
input tokens by tier reasoning short minimum
58,582 tokensFull Off suite
input tokens by tier reasoning short maximum
58,603 tokensFull Off suite
capped
0 countOriginal complete Off suite; capped and empty may overlap
empty final
0 countOriginal complete Off suite; capped and empty may overlap
overlap capped and empty
0 countOriginal complete Off suite; capped and empty may overlap
9 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
267 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.253762 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.194954 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
64.371956 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
95.204681 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.40332 GiBsampled whole host free memory; not model allocation
native decode file blobs sum
0 file blobsNative decode counters only
native decode file mb
0 MBNative decode counters only
whole run cumulative expert file blobs last
1,602 file blobsEngine total since resident copy, includes activity outside decode; file-backed access is not proof of physical SSD reads.
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
82.83
148.71
13.98
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · Compass3
unweighted correct/questions and independently reported tier-weighted correct points/possible points; primary column is tier-weighted accuracy
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · Compass3. Scores are out of 100; unweighted correct/questions and independently reported tier-weighted correct points/possible points; primary column is tier-weighted accuracy. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
70 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.47244 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.666047 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
46.715065 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
78.126431 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.673828 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
99.50
31.74
3.41
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
36 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.064851 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.123502 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
94.274258 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
133.044422 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.616211 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
99.50
88.27
1.10
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
37 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.361489 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.66405 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
79.314663 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
110.078472 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.474609 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
99.17
82.40
1.53
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
39 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062036 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.079357 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
94.532978 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
129.409496 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.186523 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
99.00
71.32
1.15
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
46 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.417591 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.504681 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
66.95652 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
98.705212 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.460938 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
98.83
49.97
2.05
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
41 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.387277 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.70464 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
70.957832 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
101.643085 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.46875 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
97.83
66.95
1.74
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
34 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.242906 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.474246 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
96.249508 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
128.280319 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.46875 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
81.83
122.13
1.41
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · Compass3X
unweighted correct/questions and independently reported tier-weighted correct points/possible points; primary column is tier-weighted accuracy
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · Compass3X. Scores are out of 100; unweighted correct/questions and independently reported tier-weighted correct points/possible points; primary column is tier-weighted accuracy. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
80 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.410193 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.139028 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
45.897488 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
77.503063 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.666992 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
95.79
32.91
3.91
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
40 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.243977 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.465271 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
80.482891 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
112.55069 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.46875 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
94.68
99.75
1.63
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
43 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.327361 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.534447 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
71.819916 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
103.161804 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.466797 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
92.47
77.51
1.80
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
39 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062346 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.107812 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
92.94389 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
131.659164 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.616211 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
91.47
85.54
1.17
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
41 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062239 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.09373 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
93.035233 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
127.925159 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.186523 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
90.47
70.28
1.23
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
50 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.353648 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.675983 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
66.160461 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
97.751095 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.458984 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
88.95
56.11
2.19
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
35 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.216868 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.413056 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
95.979527 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
127.413174 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.464844 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.63
123.33
1.42
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · Evidence-to-Action
324 components per model: value correctness, exact citation-set correctness, and their conjunction; primary column is grounded accuracy
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · Evidence-to-Action. Scores are out of 100; 324 components per model: value correctness, exact citation-set correctness, and their conjunction; primary column is grounded accuracy. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
18 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
135 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.427773 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.460264 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
44.355946 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
77.939152 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.657227 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
66.67
63.48
7.14
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
18 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
104 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.229088 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.519643 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
77.847137 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
111.164494 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.469727 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
63.58
238.00
5.22
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
18 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
118 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.288186 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.068921 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
68.910923 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
101.489025 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.454102 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
60.49
178.08
6.06
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
18 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
164 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.061851 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.134888 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
88.103401 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
122.591167 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.174805 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
58.02
65.17
5.52
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
18 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
77 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.225523 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.382021 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
94.252346 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
128.093468 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.483398 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.79
233.42
3.79
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
18 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
141 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.257787 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.809845 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
63.586113 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
96.98814 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.47168 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
53.09
151.89
7.27
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
18 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
137 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.06206 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.10772 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.676723 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
126.222351 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.606445 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
53.09
79.25
4.63
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · ShellGame4096
correct/200 final parsed state-tracking answers under 4096 output cap
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · ShellGame4096. Scores are out of 100; correct/200 final parsed state-tracking answers under 4096 output cap. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
342 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.4597 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.18259 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
63.270554 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
94.715599 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.463867 GiBsampled whole host free memory; not model allocation
style additional material count
84 countPost-hoc response-style association; not a causal reasoning effect
style additional material correct
84 countPost-hoc response-style association; not a causal reasoning effect
style additional material accuracy percent
100 %Post-hoc response-style association; not a causal reasoning effect
style bare count
116 countPost-hoc response-style association; not a causal reasoning effect
style bare correct
18 countPost-hoc response-style association; not a causal reasoning effect
style bare accuracy percent
15.517241 %Post-hoc response-style association; not a causal reasoning effect
style additional material failures
0 countPost-hoc response-style association; not a causal reasoning effect
style additional material capped failures
0 countPost-hoc response-style association; not a causal reasoning effect
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
51.00
139.01
19.09
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
248 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.440927 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
4.85473 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
70.695782 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
102.118969 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.46582 GiBsampled whole host free memory; not model allocation
style additional material count
72 countPost-hoc response-style association; not a causal reasoning effect
style additional material correct
71 countPost-hoc response-style association; not a causal reasoning effect
style additional material accuracy percent
98.611111 %Post-hoc response-style association; not a causal reasoning effect
style bare count
128 countPost-hoc response-style association; not a causal reasoning effect
style bare correct
22 countPost-hoc response-style association; not a causal reasoning effect
style bare accuracy percent
17.1875 %Post-hoc response-style association; not a causal reasoning effect
style additional material failures
1 countPost-hoc response-style association; not a causal reasoning effect
style additional material capped failures
1 countPost-hoc response-style association; not a causal reasoning effect
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
46.50
185.09
13.61
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
485 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.520426 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.411285 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
42.637905 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
73.614368 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.688477 GiBsampled whole host free memory; not model allocation
style additional material count
62 countPost-hoc response-style association; not a causal reasoning effect
style additional material correct
60 countPost-hoc response-style association; not a causal reasoning effect
style additional material accuracy percent
96.774194 %Post-hoc response-style association; not a causal reasoning effect
style bare count
138 countPost-hoc response-style association; not a causal reasoning effect
style bare correct
18 countPost-hoc response-style association; not a causal reasoning effect
style bare accuracy percent
13.043478 %Post-hoc response-style association; not a causal reasoning effect
style additional material failures
2 countPost-hoc response-style association; not a causal reasoning effect
style additional material capped failures
2 countPost-hoc response-style association; not a causal reasoning effect
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
39.00
60.59
27.75
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
194 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.06202 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.095652 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.336361 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
125.328068 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.602539 GiBsampled whole host free memory; not model allocation
style additional material count
30 countPost-hoc response-style association; not a causal reasoning effect
style additional material correct
30 countPost-hoc response-style association; not a causal reasoning effect
style additional material accuracy percent
100 %Post-hoc response-style association; not a causal reasoning effect
style bare count
170 countPost-hoc response-style association; not a causal reasoning effect
style bare correct
20 countPost-hoc response-style association; not a causal reasoning effect
style bare accuracy percent
11.764706 %Post-hoc response-style association; not a causal reasoning effect
style additional material failures
0 countPost-hoc response-style association; not a causal reasoning effect
style additional material capped failures
0 countPost-hoc response-style association; not a causal reasoning effect
paired 27b q4 correct q6 wrong
27 countSame 200 ShellGame4096 items; Q4 Uncensored vs Q6 Standard, not an isolated quantization effect
paired 27b q6 correct q4 wrong
3 countSame 200 ShellGame4096 items; Q4 Uncensored vs Q6 Standard, not an isolated quantization effect
paired 27b both correct
23 countSame 200 ShellGame4096 items; Q4 Uncensored vs Q6 Standard, not an isolated quantization effect
paired 27b both wrong
147 countSame 200 ShellGame4096 items; Q4 Uncensored vs Q6 Standard, not an isolated quantization effect
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
25.00
82.79
6.56
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
99 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.359492 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.553293 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
78.45546 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
109.853668 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.479492 GiBsampled whole host free memory; not model allocation
style additional material count
21 countPost-hoc response-style association; not a causal reasoning effect
style additional material correct
20 countPost-hoc response-style association; not a causal reasoning effect
style additional material accuracy percent
95.238095 %Post-hoc response-style association; not a causal reasoning effect
style bare count
179 countPost-hoc response-style association; not a causal reasoning effect
style bare correct
24 countPost-hoc response-style association; not a causal reasoning effect
style bare accuracy percent
13.407821 %Post-hoc response-style association; not a causal reasoning effect
style additional material failures
1 countPost-hoc response-style association; not a causal reasoning effect
style additional material capped failures
1 countPost-hoc response-style association; not a causal reasoning effect
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
22.00
217.75
5.01
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
40 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.302016 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.594766 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
93.877029 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
124.961529 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.483398 GiBsampled whole host free memory; not model allocation
style additional material count
4 countPost-hoc response-style association; not a causal reasoning effect
style additional material correct
4 countPost-hoc response-style association; not a causal reasoning effect
style additional material accuracy percent
100 %Post-hoc response-style association; not a causal reasoning effect
style bare count
196 countPost-hoc response-style association; not a causal reasoning effect
style bare correct
28 countPost-hoc response-style association; not a causal reasoning effect
style bare accuracy percent
14.285714 %Post-hoc response-style association; not a causal reasoning effect
style additional material failures
0 countPost-hoc response-style association; not a causal reasoning effect
style additional material capped failures
0 countPost-hoc response-style association; not a causal reasoning effect
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
16.00
164.61
1.82
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
59 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062173 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.091319 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.292522 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
121.504021 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.170898 GiBsampled whole host free memory; not model allocation
style additional material count
1 countPost-hoc response-style association; not a causal reasoning effect
style additional material correct
1 countPost-hoc response-style association; not a causal reasoning effect
style additional material accuracy percent
100 %Post-hoc response-style association; not a causal reasoning effect
style bare count
199 countPost-hoc response-style association; not a causal reasoning effect
style bare correct
25 countPost-hoc response-style association; not a causal reasoning effect
style bare accuracy percent
12.562814 %Post-hoc response-style association; not a causal reasoning effect
style additional material failures
0 countPost-hoc response-style association; not a causal reasoning effect
style additional material capped failures
0 countPost-hoc response-style association; not a causal reasoning effect
paired 27b q4 correct q6 wrong
27 countSame 200 ShellGame4096 items; Q4 Uncensored vs Q6 Standard, not an isolated quantization effect
paired 27b q6 correct q4 wrong
3 countSame 200 ShellGame4096 items; Q4 Uncensored vs Q6 Standard, not an isolated quantization effect
paired 27b both correct
23 countSame 200 ShellGame4096 items; Q4 Uncensored vs Q6 Standard, not an isolated quantization effect
paired 27b both wrong
147 countSame 200 ShellGame4096 items; Q4 Uncensored vs Q6 Standard, not an isolated quantization effect
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
13.00
69.55
1.86
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · Mirage · neutral
100 times the harmonic mean of real-answer accuracy and unanswerable resistance; also sensitivity excluding 48 predefined controls
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · Mirage · neutral. Scores are out of 100; 100 times the harmonic mean of real-answer accuracy and unanswerable resistance; also sensitivity excluding 48 predefined controls. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
66 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.300949 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.526795 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
63.669983 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
95.305317 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.461914 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
92.60
126.07
3.05
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
115 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.349302 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.735922 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
41.582245 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
72.08704 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.520508 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
92.60
75.15
5.75
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
55 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.31367 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
4.699504 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
69.050568 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
100.330982 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.473633 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
91.60
167.56
2.47
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
65 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062523 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.094454 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
88.021854 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
122.929237 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.174805 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
90.00
69.26
2.07
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
52 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.270183 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.369107 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
78.03603 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
110.242458 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.473633 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
89.50
194.91
2.25
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
58 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062606 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.12537 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.80579 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
126.497833 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.604492 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
83.40
84.78
1.85
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
38 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.285316 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.374741 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
93.33725 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
125.059353 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.469727 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
74.30
182.46
1.65
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · Mirage · pressure
100 times the harmonic mean of real-answer accuracy and unanswerable resistance; also sensitivity excluding 48 predefined controls
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · Mirage · pressure. Scores are out of 100; 100 times the harmonic mean of real-answer accuracy and unanswerable resistance; also sensitivity excluding 48 predefined controls. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
62 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.293352 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
4.852758 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
63.614944 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
95.170906 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.461914 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
90.60
130.94
2.86
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
55 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.305742 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.477042 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
69.826714 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
101.478966 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.477539 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
90.40
173.97
2.44
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
112 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.315899 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.442966 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
42.974812 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
74.089935 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.503906 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
90.40
75.67
5.59
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
50 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.319057 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.694226 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
76.374634 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
107.579483 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.481445 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
89.60
185.01
2.25
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
64 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062448 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.080046 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.700394 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
122.612919 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.174805 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
85.50
69.34
2.03
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
57 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062241 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.080489 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.466801 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
126.159748 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.606445 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
76.60
84.78
1.81
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score. Sensitivity excludes 48 predefined controls; this is not a fresh run.
300 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
38 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.251695 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.340194 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
94.053703 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
126.185028 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.475586 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
76.40
181.01
1.63
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · SimonSays · bank 20261010
strict all-rules pass /1000 across five 200-request banks
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · SimonSays · bank 20261010. Scores are out of 100; strict all-rules pass /1000 across five 200-request banks. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
84 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.470476 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.647862 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
67.809704 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
94.596416 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.181641 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
63.50
126.80
4.30
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
82 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.411378 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.182949 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
76.80201 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
103.567665 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.186523 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
60.00
138.40
4.16
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
173 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.472389 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
4.821093 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
41.375389 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
69.408127 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.461914 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
59.50
55.76
9.44
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
174 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062258 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.110904 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.305473 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
121.468204 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.168945 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
58.50
67.63
5.82
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
95 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.445009 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.691246 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
61.16412 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
87.881653 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.1875 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
58.00
95.38
4.96
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
155 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062011 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.093087 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.420929 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
125.817001 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.602539 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
57.00
82.95
5.22
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
62 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.350476 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.716442 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
93.957943 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
123.146454 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.21582 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
53.50
158.63
3.10
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · SimonSays · bank 1
strict all-rules pass /1000 across five 200-request banks
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · SimonSays · bank 1. Scores are out of 100; strict all-rules pass /1000 across five 200-request banks. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
184 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.475421 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
4.001928 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
40.298717 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
68.869514 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.459961 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.50
57.99
10.00
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
138 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062039 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.095558 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.734409 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
126.451988 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.604492 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.00
83.09
4.62
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
172 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062152 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.097598 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.83102 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
122.75058 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.170898 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.00
67.72
5.75
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
80 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.453599 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.199206 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
69.034782 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
96.505585 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.181641 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
54.50
128.57
4.03
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
80 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.416052 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.642975 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
75.604786 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
102.142048 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.183594 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
53.50
134.20
4.02
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
66 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.333946 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.562131 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
93.965588 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
124.551201 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.199219 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
49.50
162.97
3.24
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
99 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.440261 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.460174 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
62.192348 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
89.611107 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.182617 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
49.00
98.70
5.14
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · SimonSays · bank 2
strict all-rules pass /1000 across five 200-request banks
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · SimonSays · bank 2. Scores are out of 100; strict all-rules pass /1000 across five 200-request banks. Time basis: Summed request minutes.
Model & configuration
Score / 100
Generation tok/s
Summed request minutes
Memory see basis
Evidence
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
80 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.426355 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.518292 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
76.417393 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
103.767944 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.191406 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
59.00
132.43
4.04
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
130 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062133 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.108003 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.558414 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
126.077335 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.602539 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
57.50
83.07
4.35
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
62 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.363183 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.49276 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
93.4575 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
122.811543 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.186523 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
57.00
159.84
3.08
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
81 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.444836 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.615106 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
68.591301 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
95.873013 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.19043 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.50
126.02
4.10
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
175 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.51886 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.212548 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
40.135822 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
68.258659 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.461914 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.00
56.68
9.60
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
163 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.06226 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.093317 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
86.455284 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
120.635727 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.168945 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.00
67.76
5.45
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
97 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.45754 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.875826 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
61.71175 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
88.675781 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.199219 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
55.00
95.90
5.01
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · SimonSays · bank 3
strict all-rules pass /1000 across five 200-request banks
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · SimonSays · bank 3. Scores are out of 100; strict all-rules pass /1000 across five 200-request banks. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
79 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.440753 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.600707 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
68.767548 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
96.246113 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.176758 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
57.50
128.53
3.96
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
170 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.487218 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
5.186397 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
40.239723 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
68.68362 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.457031 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.00
57.77
9.28
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
137 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.06227 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.10675 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
88.472301 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
127.728699 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.604492 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.00
83.10
4.60
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
58 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.334286 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.487933 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
94.293247 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
124.313133 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.180664 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
55.50
159.54
2.81
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
97 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.472219 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.691877 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
60.816261 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
87.587105 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.179688 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
55.50
95.58
5.07
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
76 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.427223 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
4.449889 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
75.735832 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
102.623867 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.182617 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
54.50
132.92
3.81
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
161 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062121 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.107465 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
86.902905 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
121.154488 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.170898 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
51.50
67.72
5.37
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · SimonSays · bank 4
strict all-rules pass /1000 across five 200-request banks
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · SimonSays · bank 4. Scores are out of 100; strict all-rules pass /1000 across five 200-request banks. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
173 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.48425 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.712185 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
40.171188 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
68.636044 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.459961 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
57.50
56.50
9.39
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
78 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.395565 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.521283 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
76.378727 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
103.817253 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.188477 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
57.00
133.91
3.91
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
81 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.444087 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.598265 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
68.751266 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
96.243652 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.193359 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
57.00
127.02
4.08
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
163 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062252 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.107481 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.440701 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
121.971359 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.168945 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
57.00
67.80
5.44
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
98 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.457833 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.614156 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
61.854942 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
89.323196 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.1875 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
56.00
96.68
5.13
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
63 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.363792 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.55083 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
92.12151 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
118.245499 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.165039 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
55.50
159.39
3.13
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
200 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
141 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062306 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.108291 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
87.105579 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
125.281029 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.604492 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
51.50
82.96
4.72
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · Ops Repo Repair
solved tasks/8, assertion micro accuracy/38, and equal-weight mean of eight module accuracies
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · Ops Repo Repair. Scores are out of 100; solved tasks/8, assertion micro accuracy/38, and equal-weight mean of eight module accuracies. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
8 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
9 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.301194 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.395181 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
90.69453 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
120.108131 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.227539 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
100.00
203.92
0.18
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
8 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
15 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.320884 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.547771 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
66.989639 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
96.344807 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.225586 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
100.00
121.03
0.34
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
8 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
13 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.061761 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.077302 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
92.643482 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
131.388004 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
13.604492 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
100.00
82.61
0.36
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
8 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
16 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
2.062568 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
2.079547 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
92.935154 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
127.95351 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
9.174805 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
100.00
67.41
0.43
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
8 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
13 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.362785 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.64084 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
73.028408 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
101.457104 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.213867 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
87.50
130.75
0.33
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
8 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
26 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.449401 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.595442 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
42.832939 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
70.101604 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.467773 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
87.50
48.96
0.96
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
17,395 unique answers across 98 full Off suites. Matched Off smoke and pooled SimonSays reuse these answers. All caps and empty finals retained; no cross-benchmark score.
8 requestsTimestamp differences are reconstructed, not monotonic measurements
sampled memory sample count
15 countsampled whole host free memory; not model allocation
sampled memory median sample interval seconds
3.308046 secondssampled whole host free memory; not model allocation
sampled memory maximum sample interval seconds
3.583472 secondssampled whole host free memory; not model allocation
sampled memory minimum free ram gi b
62.619026 GiBsampled whole host free memory; not model allocation
sampled memory minimum free commit gi b
89.763874 GiBsampled whole host free memory; not model allocation
sampled memory minimum free vram gi b
2.227539 GiBsampled whole host free memory; not model allocation
Whole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
75.00
99.47
0.40
—
CompleteWhole-host memory is sampled free capacity, not model allocation. Wall times retain their measurement method; arithmetic mean decode rate is diagnostic, aggregate rate is primary.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Flash-Next full Off / Original captures October 2026; reviewed 21 October
Flash-Next Off · SimonSays pooled five banks
strict all-rules pass /1000 across five 200-request banks
Reuses the five bank captures above; 1,000 answers per model, not additional inference. Bank seed differs from inference seed5090. Caps and empty finals retained.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Off · SimonSays pooled five banks. Scores are out of 100; strict all-rules pass /1000 across five 200-request banks. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Reuses the five bank captures above; 1,000 answers per model, not additional inference. Bank seed differs from inference seed5090. Caps and empty finals retained.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
Up to 2
Bank seeds
[1, 2, 3, 4, 20261010]
Settings
See individual bank results for exact samplers and output limits
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Reuses the five bank captures above; 1,000 answers per model, not additional inference. Bank seed differs from inference seed5090. Caps and empty finals retained.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Bank seeds
[1, 2, 3, 4, 20261010]
Settings
See individual bank results for exact samplers and output limits
Pooled summary of five existing banks. Timing is summed; no pooled median inferred.
57.10
56.95
47.72
—
CompletePooled summary of five existing banks. Timing is summed; no pooled median inferred.
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Reuses the five bank captures above; 1,000 answers per model, not additional inference. Bank seed differs from inference seed5090. Caps and empty finals retained.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
Up to 2
Bank seeds
[1, 2, 3, 4, 20261010]
Settings
See individual bank results for exact samplers and output limits
Pooled summary of five existing banks. Timing is summed; no pooled median inferred.
56.80
134.44
19.94
—
CompletePooled summary of five existing banks. Timing is summed; no pooled median inferred.
27B StandardOff · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Reuses the five bank captures above; 1,000 answers per model, not additional inference. Bank seed differs from inference seed5090. Caps and empty finals retained.
Thinking
Off
K / V cache
Q8_0 / Q8_0
MTP draft tokens
0
Bank seeds
[1, 2, 3, 4, 20261010]
Settings
See individual bank results for exact samplers and output limits
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Reuses the five bank captures above; 1,000 answers per model, not additional inference. Bank seed differs from inference seed5090. Caps and empty finals retained.
Thinking
Off
K / V cache
Q8_0 / Q8_0
MTP draft tokens
0
Bank seeds
[1, 2, 3, 4, 20261010]
Settings
See individual bank results for exact samplers and output limits
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Reuses the five bank captures above; 1,000 answers per model, not additional inference. Bank seed differs from inference seed5090. Caps and empty finals retained.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
Up to 2
Bank seeds
[1, 2, 3, 4, 20261010]
Settings
See individual bank results for exact samplers and output limits
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Reuses the five bank captures above; 1,000 answers per model, not additional inference. Bank seed differs from inference seed5090. Caps and empty finals retained.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
Up to 2
Bank seeds
[1, 2, 3, 4, 20261010]
Settings
See individual bank results for exact samplers and output limits
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Compass3X 60 · pairedOff. Scores are out of 100; Tier-weighted percentage on 285 points; correct-answer count and unweighted percentage separate. Time basis: Summed request minutes.
Model & configuration
Score / 100
Generation tok/s
Summed request minutes
Memory see basis
Evidence
SwiftOff (matched subset) · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
13.878339 secondsSelected requests; excludes loading, qualification, gaps and grading
questions per minute
259.39704 requests/minuteSelected requests; excludes loading, qualification, gaps and grading
completion tokens
261 tokensSelected requests; excludes loading, qualification, gaps and grading
input tokens
3,820 tokensSelected requests; excludes loading, qualification, gaps and grading
reasoning tokens
Unavailable Selected requests; excludes loading, qualification, gaps and grading
native decode seconds
3.030675 secondsSelected requests; excludes loading, qualification, gaps and grading
native decode tokens per second
86.119429 tokens/sSelected requests; excludes loading, qualification, gaps and grading
median request seconds
0.229025 secondsSelected requests; excludes loading, qualification, gaps and grading
median first content seconds
0.177216 secondsSelected requests; excludes loading, qualification, gaps and grading
median first reasoning seconds
Unavailable Selected requests; excludes loading, qualification, gaps and grading
observed whole host minimum ram gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
observed whole host minimum commit gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
observed whole host minimum vram gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
new generations
0 responsesPaired Off reuses the existing archive
Memory SOURCE_RETAINED. All outcomes retained; separate reasoning-token counts unavailable.
91.58
86.12
0.23
—
CompleteMemory SOURCE_RETAINED. All outcomes retained; separate reasoning-token counts unavailable.
27B StandardOff (matched subset) · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Evidence-to-Action 18 · pairedOff. Scores are out of 100; Grounded percentage over 324 components; semantic and citation metrics separate. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
428.647604 secondsSelected requests; excludes loading, qualification, gaps and grading
questions per minute
2.519552 requests/minuteSelected requests; excludes loading, qualification, gaps and grading
completion tokens
16,531 tokensSelected requests; excludes loading, qualification, gaps and grading
input tokens
220,092 tokensSelected requests; excludes loading, qualification, gaps and grading
reasoning tokens
Unavailable Selected requests; excludes loading, qualification, gaps and grading
native decode seconds
260.3964 secondsSelected requests; excludes loading, qualification, gaps and grading
native decode tokens per second
63.483981 tokens/sSelected requests; excludes loading, qualification, gaps and grading
median request seconds
23.627421 secondsSelected requests; excludes loading, qualification, gaps and grading
median first content seconds
9.330749 secondsSelected requests; excludes loading, qualification, gaps and grading
median first reasoning seconds
Unavailable Selected requests; excludes loading, qualification, gaps and grading
observed whole host minimum ram gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
observed whole host minimum commit gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
observed whole host minimum vram gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
new generations
0 responsesPaired Off reuses the existing archive
Memory SOURCE_RETAINED. All outcomes retained; separate reasoning-token counts unavailable.
66.67
63.48
7.14
—
CompleteMemory SOURCE_RETAINED. All outcomes retained; separate reasoning-token counts unavailable.
SwiftOff (matched subset) · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
363.564125 secondsSelected requests; excludes loading, qualification, gaps and grading
questions per minute
2.97059 requests/minuteSelected requests; excludes loading, qualification, gaps and grading
completion tokens
16,337 tokensSelected requests; excludes loading, qualification, gaps and grading
input tokens
220,092 tokensSelected requests; excludes loading, qualification, gaps and grading
reasoning tokens
Unavailable Selected requests; excludes loading, qualification, gaps and grading
native decode seconds
91.7402 secondsSelected requests; excludes loading, qualification, gaps and grading
native decode tokens per second
178.078966 tokens/sSelected requests; excludes loading, qualification, gaps and grading
median request seconds
19.942582 secondsSelected requests; excludes loading, qualification, gaps and grading
median first content seconds
14.75713 secondsSelected requests; excludes loading, qualification, gaps and grading
median first reasoning seconds
Unavailable Selected requests; excludes loading, qualification, gaps and grading
observed whole host minimum ram gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
observed whole host minimum commit gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
observed whole host minimum vram gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
new generations
0 responsesPaired Off reuses the existing archive
Memory SOURCE_RETAINED. All outcomes retained; separate reasoning-token counts unavailable.
60.49
178.08
6.06
—
CompleteMemory SOURCE_RETAINED. All outcomes retained; separate reasoning-token counts unavailable.
27B StandardOff (matched subset) · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next ShellGame 40 · pairedOff. Scores are out of 100; Correct state-tracking answers /40, expressed as percentage. Time basis: Summed request minutes.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
75.595558 secondsSelected requests; excludes loading, qualification, gaps and grading
questions per minute
31.747897 requests/minuteSelected requests; excludes loading, qualification, gaps and grading
completion tokens
4,969 tokensSelected requests; excludes loading, qualification, gaps and grading
input tokens
16,431 tokensSelected requests; excludes loading, qualification, gaps and grading
reasoning tokens
Unavailable Selected requests; excludes loading, qualification, gaps and grading
native decode seconds
60.023745 secondsSelected requests; excludes loading, qualification, gaps and grading
native decode tokens per second
82.783905 tokens/sSelected requests; excludes loading, qualification, gaps and grading
median request seconds
0.523801 secondsSelected requests; excludes loading, qualification, gaps and grading
median first content seconds
0.398204 secondsSelected requests; excludes loading, qualification, gaps and grading
median first reasoning seconds
Unavailable Selected requests; excludes loading, qualification, gaps and grading
observed whole host minimum ram gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
observed whole host minimum commit gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
observed whole host minimum vram gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
new generations
0 responsesPaired Off reuses the existing archive
Memory SOURCE_RETAINED. All outcomes retained; separate reasoning-token counts unavailable.
32.50
82.78
1.26
—
CompleteMemory SOURCE_RETAINED. All outcomes retained; separate reasoning-token counts unavailable.
SwiftOff (matched subset) · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
21.158919 secondsSelected requests; excludes loading, qualification, gaps and grading
questions per minute
113.427343 requests/minuteSelected requests; excludes loading, qualification, gaps and grading
completion tokens
544 tokensSelected requests; excludes loading, qualification, gaps and grading
input tokens
16,431 tokensSelected requests; excludes loading, qualification, gaps and grading
reasoning tokens
Unavailable Selected requests; excludes loading, qualification, gaps and grading
native decode seconds
3.5922 secondsSelected requests; excludes loading, qualification, gaps and grading
native decode tokens per second
151.439229 tokens/sSelected requests; excludes loading, qualification, gaps and grading
median request seconds
0.445982 secondsSelected requests; excludes loading, qualification, gaps and grading
median first content seconds
0.366438 secondsSelected requests; excludes loading, qualification, gaps and grading
median first reasoning seconds
Unavailable Selected requests; excludes loading, qualification, gaps and grading
observed whole host minimum ram gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
observed whole host minimum commit gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
observed whole host minimum vram gi b
Unavailable Sampled whole-host free memory; not process allocation. SOURCE_RETAINED
new generations
0 responsesPaired Off reuses the existing archive
Memory SOURCE_RETAINED. All outcomes retained; separate reasoning-token counts unavailable.
15.00
151.44
0.35
—
CompleteMemory SOURCE_RETAINED. All outcomes retained; separate reasoning-token counts unavailable.
27B StandardOff (matched subset) · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Compass3X 60 · xHigh. Scores are out of 100; Tier-weighted percentage on 285 points; correct-answer count and unweighted percentage separate. Time basis: Summed request minutes.
Model & configuration
Score / 100
Generation tok/s
Summed request minutes
Memory see basis
Evidence
SwiftxHigh · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
69.424724 secondsSelected requests; excludes loading, qualification, gaps and grading
questions per minute
51.854726 requests/minuteSelected requests; excludes loading, qualification, gaps and grading
completion tokens
4,663 tokensSelected requests; excludes loading, qualification, gaps and grading
input tokens
5,980 tokensSelected requests; excludes loading, qualification, gaps and grading
reasoning tokens
Unavailable Selected requests; excludes loading, qualification, gaps and grading
native decode seconds
55.866054 secondsSelected requests; excludes loading, qualification, gaps and grading
native decode tokens per second
83.467502 tokens/sSelected requests; excludes loading, qualification, gaps and grading
median request seconds
0.918591 secondsSelected requests; excludes loading, qualification, gaps and grading
median first content seconds
0.869163 secondsSelected requests; excludes loading, qualification, gaps and grading
median first reasoning seconds
0.20822 secondsSelected requests; excludes loading, qualification, gaps and grading
observed whole host minimum ram gi b
86.220592 GiBSampled whole-host free memory; not process allocation. AVAILABLE
observed whole host minimum commit gi b
126.166962 GiBSampled whole-host free memory; not process allocation. AVAILABLE
observed whole host minimum vram gi b
13.660156 GiBSampled whole-host free memory; not process allocation. AVAILABLE
new generations
60 responsesPaired Off reuses the existing archive
Memory AVAILABLE. All outcomes retained; separate reasoning-token counts unavailable.
100.00
83.47
1.16
—
CompleteMemory AVAILABLE. All outcomes retained; separate reasoning-token counts unavailable.
27B StandardxHigh · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next Evidence-to-Action 18 · xHigh. Scores are out of 100; Grounded percentage over 324 components; semantic and citation metrics separate. Time basis: Summed request minutes.
Model & configuration
Score / 100
Generation tok/s
Summed request minutes
Memory see basis
Evidence
27B StandardxHigh · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
925.87373 secondsSelected requests; excludes loading, qualification, gaps and grading
questions per minute
1.166466 requests/minuteSelected requests; excludes loading, qualification, gaps and grading
completion tokens
134,120 tokensSelected requests; excludes loading, qualification, gaps and grading
input tokens
220,740 tokensSelected requests; excludes loading, qualification, gaps and grading
reasoning tokens
Unavailable Selected requests; excludes loading, qualification, gaps and grading
native decode seconds
759.6541 secondsSelected requests; excludes loading, qualification, gaps and grading
native decode tokens per second
176.55404 tokens/sSelected requests; excludes loading, qualification, gaps and grading
median request seconds
52.314529 secondsSelected requests; excludes loading, qualification, gaps and grading
median first content seconds
48.700618 secondsSelected requests; excludes loading, qualification, gaps and grading
median first reasoning seconds
8.989594 secondsSelected requests; excludes loading, qualification, gaps and grading
observed whole host minimum ram gi b
85.233112 GiBSampled whole-host free memory; not process allocation. AVAILABLE
observed whole host minimum commit gi b
114.403702 GiBSampled whole-host free memory; not process allocation. AVAILABLE
observed whole host minimum vram gi b
2.213867 GiBSampled whole-host free memory; not process allocation. AVAILABLE
new generations
18 responsesPaired Off reuses the existing archive
Memory AVAILABLE. All outcomes retained; separate reasoning-token counts unavailable.
94.44
176.55
15.43
—
CompleteMemory AVAILABLE. All outcomes retained; separate reasoning-token counts unavailable.
SwiftxHigh · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
7 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next ShellGame 40 · xHigh. Scores are out of 100; Correct state-tracking answers /40, expressed as percentage. Time basis: Summed request minutes.
Model & configuration
Score / 100
Generation tok/s
Summed request minutes
Memory see basis
Evidence
SwiftxHigh · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
525.51713 secondsSelected requests; excludes loading, qualification, gaps and grading
questions per minute
4.56693 requests/minuteSelected requests; excludes loading, qualification, gaps and grading
completion tokens
41,746 tokensSelected requests; excludes loading, qualification, gaps and grading
input tokens
17,871 tokensSelected requests; excludes loading, qualification, gaps and grading
reasoning tokens
Unavailable Selected requests; excludes loading, qualification, gaps and grading
native decode seconds
509.186609 secondsSelected requests; excludes loading, qualification, gaps and grading
native decode tokens per second
81.98566 tokens/sSelected requests; excludes loading, qualification, gaps and grading
median request seconds
9.136087 secondsSelected requests; excludes loading, qualification, gaps and grading
median first content seconds
9.066746 secondsSelected requests; excludes loading, qualification, gaps and grading
median first reasoning seconds
0.39448 secondsSelected requests; excludes loading, qualification, gaps and grading
observed whole host minimum ram gi b
84.686897 GiBSampled whole-host free memory; not process allocation. AVAILABLE
observed whole host minimum commit gi b
124.321018 GiBSampled whole-host free memory; not process allocation. AVAILABLE
observed whole host minimum vram gi b
13.438477 GiBSampled whole-host free memory; not process allocation. AVAILABLE
new generations
40 responsesPaired Off reuses the existing archive
Memory AVAILABLE. All outcomes retained; separate reasoning-token counts unavailable.
97.50
81.99
8.76
—
CompleteMemory AVAILABLE. All outcomes retained; separate reasoning-token counts unavailable.
27B StandardxHigh · Q8_0 / Q8_0Captured: Capture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
Selected request start/finish bounds; includes gaps, not summed duration
Capture started
Capture finished
Capture timezone
UTC
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Final-amendment4. 826 matched Off answers across all three profiles are a subset of the 17,395 Off archive; not new generations. 826 xHigh answers are a separate cohort. Caps and empty finals retained. Sampler and output cap differ.
2026-10-12 to 2026-10-12 UTC; per-result timestamps belowStage or answer dates in result details
Application / version
Strata native · version not recorded
Runtime / extension
Strata 0.1.40
Benchmark
HH2-SCORE-1.0
Context capacity
262,144 tokens loaded
Thinking
Off
MTP
Per model and cohort; see recorded settings
K / V cache
Strata int8 or Prism Q8 K/V; see each result
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
5 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next HH2 · SHORT input condition. Scores are out of 100; HH2 composite:40% present +20% per-haystack floored signed absent credit +40% reasoning; mean across12 haystacks. Time basis: Summed request minutes.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
2026-10-12T03:43:05.536707+00:00; derived from duration, not an independent timestamp
HH2 method
{"unitOfAnalysis": "generated haystack", "haystacks": 12, "composite": "Mean over 12 haystacks of 100 * (0.4 * presentCorrect/80 + 0.2 * max(0, absentSignedCredit/24) + 0.4 * endToEndReasoningCorrect/20)", "absentCredit": "+1 recognized missing-value response; -1 recognized fabricated numeric value; 0 other failure; floor within each haystack", "reasoningCredit": "Joint ordered value retrieval, binding, arithmetic, and accepted output syntax", "reasoningFailureClassCaveat": "Recorded parser classes describe the frozen grammar; an unresolved whole line does not prove its visible values or arithmetic were wrong. Consult the separate unsupported-syntax count.", "notACombinedCrossBenchmarkScore": true, "timingScope": "Sum of saved request durations; loading, qualification (for primary totals), grading, cleanup, gaps, earlier failed work excluded", "versions": {"aggregator": "HH2-AGG-1.1", "benchmark": "HH-2.0", "classic_parser": "HH2-LENIENT-1.0", "generator": "HH2-GEN-1.0", "manifest": "HH2-MANIFEST-1.1", "prompt": "HH2-PROMPT-1.1", "reasoning_parser": "SEM-PC-1 / PROTO-PIPE-1", "runner": "HH2-RUN-1.1", "scorer": "HH2-SCORE-1.0", "strict_parser": "HH2-STRICT-1.0", "think_strip": "HH2-THINK-1.0"}, "statisticalFamily": {"tests": 20, "endpoints": ["accuracy", "retention"], "alpha": 0.05, "holmTests": true, "bonferroniIntervalTails": true, "practicalMarginPoints": 2, "bootstrapSeed": 5090, "bootstrapReplicates": 100000}}
All measurements and denominators (45)
haystacks
12 countHH2 SHORT; signed absent credit may be negative before per-haystack floor
primary requests
24 countHH2 SHORT; signed absent credit may be negative before per-haystack floor
composite score out of100
99.694444 points /100HH2 SHORT; signed absent credit may be negative before per-haystack floor
present correct
960 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
present denominator
960 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
recognized absent correct
287 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
recognized numeric fabrications
1 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
absent denominator
288 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
absent other failures
0 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
absent signed credit
286 credit pointsHH2 SHORT; signed absent credit may be negative before per-haystack floor
absent credit after per haystack floor
286 credit pointsHH2 SHORT; signed absent credit may be negative before per-haystack floor
end to end reasoning correct
239 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
reasoning denominator
240 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
reasoning misses
1 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
classic strict syntax compliant
1,248 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
classic syntax denominator
1,248 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
reasoning pipe syntax compliant
240 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
reasoning syntax denominator
240 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
reasoning misses with correct visible numbers unsupported syntax
0 componentsHH2 SHORT; signed absent credit may be negative before per-haystack floor
reasoning sum correct
60 componentsStrict ordered binding, arithmetic and syntax
reasoning sum denominator
60 componentsStrict ordered binding, arithmetic and syntax
reasoning min correct
60 componentsStrict ordered binding, arithmetic and syntax
reasoning min denominator
60 componentsStrict ordered binding, arithmetic and syntax
reasoning max correct
60 componentsStrict ordered binding, arithmetic and syntax
reasoning max denominator
60 componentsStrict ordered binding, arithmetic and syntax
reasoning range correct
59 componentsStrict ordered binding, arithmetic and syntax
reasoning range denominator
60 componentsStrict ordered binding, arithmetic and syntax
requests
24 requestsHH2 SHORT primary requests; excludes qualification and loading
request wall seconds
343.609132 secondsHH2 SHORT primary requests; excludes qualification and loading
native prefill seconds
122.8585 secondsHH2 SHORT primary requests; excludes qualification and loading
native decode seconds
219.7795 secondsHH2 SHORT primary requests; excludes qualification and loading
prompt tokens
178,055 tokensHH2 SHORT primary requests; excludes qualification and loading
prompt read tokens
98,052 tokensHH2 SHORT primary requests; excludes qualification and loading
reused prompt tokens
80,003 tokensHH2 SHORT primary requests; excludes qualification and loading
requests with prompt reuse
12 requestsHH2 SHORT primary requests; excludes qualification and loading
generated tokens
22,368 tokensHH2 SHORT primary requests; excludes qualification and loading
median request wall seconds
13.96787 secondsHH2 SHORT primary requests; excludes qualification and loading
median time to first token seconds
4.982093 secondsHH2 SHORT primary requests; excludes qualification and loading
capped requests
0 requestsHH2 SHORT primary requests; excludes qualification and loading
empty final requests
0 requestsHH2 SHORT primary requests; excludes qualification and loading
native draft tokens offered
0 tokensHH2 SHORT primary requests; excludes qualification and loading
native draft tokens accepted
0 tokensHH2 SHORT primary requests; excludes qualification and loading
native aggregate decode tokens per second
101.774733 tokens/sHH2 SHORT primary requests; excludes qualification and loading
minimum input tokens
7,355 tokensHH2 SHORT primary requests; excludes qualification and loading
maximum input tokens
7,484 tokensHH2 SHORT primary requests; excludes qualification and loading
No statistically separated winner. Composite includes strict output syntax. Original completed cohort
99.69
101.77
5.73
—
Descriptive · qualification limitsNo statistically separated winner. Composite includes strict output syntax. Original completed cohort
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
2026-10-12 to 2026-10-12 UTC; per-result timestamps belowStage or answer dates in result details
Application / version
Strata native · version not recorded
Runtime / extension
Strata 0.1.40
Benchmark
HH2-SCORE-1.0
Context capacity
262,144 tokens loaded
Thinking
Off
MTP
Per model and cohort; see recorded settings
K / V cache
Strata int8 or Prism Q8 K/V; see each result
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
5 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next HH2 · 64K input condition. Scores are out of 100; HH2 composite:40% present +20% per-haystack floored signed absent credit +40% reasoning; mean across12 haystacks. Time basis: Summed request minutes.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
2026-10-12T03:42:37.200121+00:00; derived from duration, not an independent timestamp
HH2 method
{"unitOfAnalysis": "generated haystack", "haystacks": 12, "composite": "Mean over 12 haystacks of 100 * (0.4 * presentCorrect/80 + 0.2 * max(0, absentSignedCredit/24) + 0.4 * endToEndReasoningCorrect/20)", "absentCredit": "+1 recognized missing-value response; -1 recognized fabricated numeric value; 0 other failure; floor within each haystack", "reasoningCredit": "Joint ordered value retrieval, binding, arithmetic, and accepted output syntax", "reasoningFailureClassCaveat": "Recorded parser classes describe the frozen grammar; an unresolved whole line does not prove its visible values or arithmetic were wrong. Consult the separate unsupported-syntax count.", "notACombinedCrossBenchmarkScore": true, "timingScope": "Sum of saved request durations; loading, qualification (for primary totals), grading, cleanup, gaps, earlier failed work excluded", "versions": {"aggregator": "HH2-AGG-1.1", "benchmark": "HH-2.0", "classic_parser": "HH2-LENIENT-1.0", "generator": "HH2-GEN-1.0", "manifest": "HH2-MANIFEST-1.1", "prompt": "HH2-PROMPT-1.1", "reasoning_parser": "SEM-PC-1 / PROTO-PIPE-1", "runner": "HH2-RUN-1.1", "scorer": "HH2-SCORE-1.0", "strict_parser": "HH2-STRICT-1.0", "think_strip": "HH2-THINK-1.0"}, "statisticalFamily": {"tests": 20, "endpoints": ["accuracy", "retention"], "alpha": 0.05, "holmTests": true, "bonferroniIntervalTails": true, "practicalMarginPoints": 2, "bootstrapSeed": 5090, "bootstrapReplicates": 100000}}
All measurements and denominators (45)
haystacks
12 countHH2 64K; signed absent credit may be negative before per-haystack floor
primary requests
24 countHH2 64K; signed absent credit may be negative before per-haystack floor
composite score out of100
98.458333 points /100HH2 64K; signed absent credit may be negative before per-haystack floor
present correct
959 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
present denominator
960 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
recognized absent correct
282 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
recognized numeric fabrications
6 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
absent denominator
288 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
absent other failures
0 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
absent signed credit
276 credit pointsHH2 64K; signed absent credit may be negative before per-haystack floor
absent credit after per haystack floor
276 credit pointsHH2 64K; signed absent credit may be negative before per-haystack floor
end to end reasoning correct
236 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
reasoning denominator
240 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
reasoning misses
4 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
classic strict syntax compliant
1,248 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
classic syntax denominator
1,248 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
reasoning pipe syntax compliant
240 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
reasoning syntax denominator
240 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
reasoning misses with correct visible numbers unsupported syntax
0 componentsHH2 64K; signed absent credit may be negative before per-haystack floor
reasoning sum correct
59 componentsStrict ordered binding, arithmetic and syntax
reasoning sum denominator
60 componentsStrict ordered binding, arithmetic and syntax
reasoning min correct
60 componentsStrict ordered binding, arithmetic and syntax
reasoning min denominator
60 componentsStrict ordered binding, arithmetic and syntax
reasoning max correct
60 componentsStrict ordered binding, arithmetic and syntax
reasoning max denominator
60 componentsStrict ordered binding, arithmetic and syntax
reasoning range correct
57 componentsStrict ordered binding, arithmetic and syntax
reasoning range denominator
60 componentsStrict ordered binding, arithmetic and syntax
requests
24 requestsHH2 64K primary requests; excludes qualification and loading
request wall seconds
931.714873 secondsHH2 64K primary requests; excludes qualification and loading
native prefill seconds
712.9505 secondsHH2 64K primary requests; excludes qualification and loading
native decode seconds
216.9347 secondsHH2 64K primary requests; excludes qualification and loading
prompt tokens
1,057,395 tokensHH2 64K primary requests; excludes qualification and loading
prompt read tokens
537,722 tokensHH2 64K primary requests; excludes qualification and loading
reused prompt tokens
519,673 tokensHH2 64K primary requests; excludes qualification and loading
requests with prompt reuse
12 requestsHH2 64K primary requests; excludes qualification and loading
generated tokens
22,383 tokensHH2 64K primary requests; excludes qualification and loading
median request wall seconds
38.892483 secondsHH2 64K primary requests; excludes qualification and loading
median time to first token seconds
29.177055 secondsHH2 64K primary requests; excludes qualification and loading
capped requests
0 requestsHH2 64K primary requests; excludes qualification and loading
empty final requests
0 requestsHH2 64K primary requests; excludes qualification and loading
native draft tokens offered
0 tokensHH2 64K primary requests; excludes qualification and loading
native draft tokens accepted
0 tokensHH2 64K primary requests; excludes qualification and loading
native aggregate decode tokens per second
103.178514 tokens/sHH2 64K primary requests; excludes qualification and loading
minimum input tokens
43,964 tokensHH2 64K primary requests; excludes qualification and loading
maximum input tokens
44,142 tokensHH2 64K primary requests; excludes qualification and loading
No statistically separated winner. Composite includes strict output syntax. Original completed cohort
98.46
103.18
15.53
—
Descriptive · qualification limitsNo statistically separated winner. Composite includes strict output syntax. Original completed cohort
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
2026-10-12 to 2026-10-12 UTC; per-result timestamps belowStage or answer dates in result details
Application / version
Strata native · version not recorded
Runtime / extension
Strata 0.1.40
Benchmark
HH2-SCORE-1.0
Context capacity
262,144 tokens loaded
Thinking
Off
MTP
Per model and cohort; see recorded settings
K / V cache
Strata int8 or Prism Q8 K/V; see each result
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
5 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next HH2 · 128K input condition. Scores are out of 100; HH2 composite:40% present +20% per-haystack floored signed absent credit +40% reasoning; mean across12 haystacks. Time basis: Summed request minutes.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
2026-10-12T03:41:18.885188+00:00; derived from duration, not an independent timestamp
HH2 method
{"unitOfAnalysis": "generated haystack", "haystacks": 12, "composite": "Mean over 12 haystacks of 100 * (0.4 * presentCorrect/80 + 0.2 * max(0, absentSignedCredit/24) + 0.4 * endToEndReasoningCorrect/20)", "absentCredit": "+1 recognized missing-value response; -1 recognized fabricated numeric value; 0 other failure; floor within each haystack", "reasoningCredit": "Joint ordered value retrieval, binding, arithmetic, and accepted output syntax", "reasoningFailureClassCaveat": "Recorded parser classes describe the frozen grammar; an unresolved whole line does not prove its visible values or arithmetic were wrong. Consult the separate unsupported-syntax count.", "notACombinedCrossBenchmarkScore": true, "timingScope": "Sum of saved request durations; loading, qualification (for primary totals), grading, cleanup, gaps, earlier failed work excluded", "versions": {"aggregator": "HH2-AGG-1.1", "benchmark": "HH-2.0", "classic_parser": "HH2-LENIENT-1.0", "generator": "HH2-GEN-1.0", "manifest": "HH2-MANIFEST-1.1", "prompt": "HH2-PROMPT-1.1", "reasoning_parser": "SEM-PC-1 / PROTO-PIPE-1", "runner": "HH2-RUN-1.1", "scorer": "HH2-SCORE-1.0", "strict_parser": "HH2-STRICT-1.0", "think_strip": "HH2-THINK-1.0"}, "statisticalFamily": {"tests": 20, "endpoints": ["accuracy", "retention"], "alpha": 0.05, "holmTests": true, "bonferroniIntervalTails": true, "practicalMarginPoints": 2, "bootstrapSeed": 5090, "bootstrapReplicates": 100000}}
All measurements and denominators (45)
haystacks
12 countHH2 128K; signed absent credit may be negative before per-haystack floor
primary requests
24 countHH2 128K; signed absent credit may be negative before per-haystack floor
composite score out of100
96.694444 points /100HH2 128K; signed absent credit may be negative before per-haystack floor
present correct
958 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
present denominator
960 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
recognized absent correct
278 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
recognized numeric fabrications
10 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
absent denominator
288 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
absent other failures
0 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
absent signed credit
268 credit pointsHH2 128K; signed absent credit may be negative before per-haystack floor
absent credit after per haystack floor
268 credit pointsHH2 128K; signed absent credit may be negative before per-haystack floor
end to end reasoning correct
229 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
reasoning denominator
240 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
reasoning misses
11 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
classic strict syntax compliant
1,248 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
classic syntax denominator
1,248 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
reasoning pipe syntax compliant
240 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
reasoning syntax denominator
240 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
reasoning misses with correct visible numbers unsupported syntax
0 componentsHH2 128K; signed absent credit may be negative before per-haystack floor
reasoning sum correct
58 componentsStrict ordered binding, arithmetic and syntax
reasoning sum denominator
60 componentsStrict ordered binding, arithmetic and syntax
reasoning min correct
60 componentsStrict ordered binding, arithmetic and syntax
reasoning min denominator
60 componentsStrict ordered binding, arithmetic and syntax
reasoning max correct
60 componentsStrict ordered binding, arithmetic and syntax
reasoning max denominator
60 componentsStrict ordered binding, arithmetic and syntax
reasoning range correct
51 componentsStrict ordered binding, arithmetic and syntax
reasoning range denominator
60 componentsStrict ordered binding, arithmetic and syntax
requests
24 requestsHH2 128K primary requests; excludes qualification and loading
request wall seconds
2,041.577802 secondsHH2 128K primary requests; excludes qualification and loading
native prefill seconds
1,826.7109 secondsHH2 128K primary requests; excludes qualification and loading
native decode seconds
211.7358 secondsHH2 128K primary requests; excludes qualification and loading
prompt tokens
2,500,217 tokensHH2 128K primary requests; excludes qualification and loading
prompt read tokens
1,259,133 tokensHH2 128K primary requests; excludes qualification and loading
reused prompt tokens
1,241,084 tokensHH2 128K primary requests; excludes qualification and loading
requests with prompt reuse
12 requestsHH2 128K primary requests; excludes qualification and loading
generated tokens
22,395 tokensHH2 128K primary requests; excludes qualification and loading
median request wall seconds
82.942083 secondsHH2 128K primary requests; excludes qualification and loading
median time to first token seconds
74.141829 secondsHH2 128K primary requests; excludes qualification and loading
capped requests
0 requestsHH2 128K primary requests; excludes qualification and loading
empty final requests
0 requestsHH2 128K primary requests; excludes qualification and loading
native draft tokens offered
0 tokensHH2 128K primary requests; excludes qualification and loading
native draft tokens accepted
0 tokensHH2 128K primary requests; excludes qualification and loading
native aggregate decode tokens per second
105.768604 tokens/sHH2 128K primary requests; excludes qualification and loading
minimum input tokens
104,067 tokensHH2 128K primary requests; excludes qualification and loading
maximum input tokens
104,249 tokensHH2 128K primary requests; excludes qualification and loading
No statistically separated winner. Composite includes strict output syntax. Original completed cohort
96.69
105.77
34.03
—
Descriptive · qualification limitsNo statistically separated winner. Composite includes strict output syntax. Original completed cohort
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
2026-10-12 to 2026-10-12 UTC; per-result timestamps belowStage or answer dates in result details
Application / version
Strata native · version not recorded
Runtime / extension
Strata 0.1.40
Benchmark
HH2-SCORE-1.0
Context capacity
262,144 tokens loaded
Thinking
Off
MTP
Per model and cohort; see recorded settings
K / V cache
Strata int8 or Prism Q8 K/V; see each result
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
5 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next HH2 · 256K input condition. Scores are out of 100; HH2 composite:40% present +20% per-haystack floored signed absent credit +40% reasoning; mean across12 haystacks. Time basis: Summed request minutes.
Model & configuration
Score / 100
Generation tok/s
Summed request minutes
Memory see basis
Evidence
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
2026-10-12 to 2026-10-12 UTC; per-result timestamps belowStage or answer dates in result details
Application / version
Strata native · version not recorded
Runtime / extension
Strata 0.1.40
Benchmark
HH2 primary timing
Context capacity
262,144 tokens loaded
Thinking
Off
MTP
Per model and cohort; see recorded settings
K / V cache
Strata int8 or Prism Q8 K/V; see each result
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
5 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next HH2 · all primary requests. Scores are out of 100; No pooled quality score; separate timing and qualification diagnostics. Time basis: Summed request minutes.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
-14.5 composite pointsDifference between the 256K-labelled and SHORT primary condition composites; same 12 haystacks
Separate complete recovery; original incomplete cohort excluded
—
117.27
69.06
—
Descriptive · qualification limitsSeparate complete recovery; original incomplete cohort excluded
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
2026-10-12 to 2026-10-12 UTC; per-result timestamps belowStage or answer dates in result details
Application / version
Strata native · version not recorded
Runtime / extension
Strata 0.1.40
Benchmark
HH2 qualification
Context capacity
262,144 tokens loaded
Thinking
Off
MTP
Per model and cohort; see recorded settings
K / V cache
Strata int8 or Prism Q8 K/V; see each result
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
5 results
Aggregate native decode tokens / native decode seconds; excludes prefill and loading · — = not available
Flash-Next HH2 · qualification repeats. Scores are out of 100; No pooled quality score; separate timing and qualification diagnostics. Time basis: Summed request minutes.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
4 triplesFour selected checks; not proof of universal determinism
Separate complete recovery; original incomplete cohort excluded
—
105.28
12.63
—
Descriptive · qualification limitsSeparate complete recovery; original incomplete cohort excluded
SwiftOff · int8 / int8Captured: Capture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Date precision
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Request start bound; derived completion estimate is not an independently recorded completion timestamp
Capture started
Capture timezone
UTC
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery.
Not recordedSource date above is not a capture date
Application / version
Strata native · version not recorded
Runtime / extension
Strata 0.1.40
Benchmark
HH2 paired statistics
Context capacity
262,144 tokens loaded
Thinking
Off
MTP
Per model and cohort; see recorded settings
K / V cache
Strata int8 or Prism Q8 K/V; see each result
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
10 results
Not applicable to statistical summaries · — = not available
Flash-Next HH2 · paired accuracy uncertainty. Scores are out of 100; No pooled quality score; signed model A minus model B effect with multiplicity-aware uncertainty. Time basis: Summed request minutes.
Model & configuration
Score / 100
Generation tok/s
Summed request minutes
Memory see basis
Evidence
Coder vs ISTAOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Coder Qwen3.8-Flash-Next IQ1_M versus ISTA Qwen3.8-Flash-Next IQ3_S
ISTA Coder Qwen3.8-Flash-Next IQ1_M versus ISTA Qwen3.8-Flash-Next IQ3_S
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Coder; model B: ISTA
Scope
Derived from primary HH2 captures; not another run
-2.5 composite pointsPaired accuracy endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.567871 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-8.486111 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
10.375017 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
Coder vs SwiftOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Coder Qwen3.8-Flash-Next IQ1_M versus Swift Qwen3.8-Flash-Next IQ3_XXS
ISTA Coder Qwen3.8-Flash-Next IQ1_M versus Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Coder; model B: Swift
Scope
Derived from primary HH2 captures; not another run
-6.180556 composite pointsPaired accuracy endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.213867 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-18.472292 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
7.263889 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
Coder vs OrcaRouterOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Coder Qwen3.8-Flash-Next IQ1_M versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
ISTA Coder Qwen3.8-Flash-Next IQ1_M versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Coder; model B: OrcaRouter
Scope
Derived from primary HH2 captures; not another run
3.125 composite pointsPaired accuracy endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.483887 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-10.319444 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
14.375 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
Coder vs Flash-Next UnslothOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Coder Qwen3.8-Flash-Next IQ1_M versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
ISTA Coder Qwen3.8-Flash-Next IQ1_M versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Coder; model B: Flash-Next Unsloth
Scope
Derived from primary HH2 captures; not another run
-2.986111 composite pointsPaired accuracy endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.590332 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-17.972222 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
12.111111 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
ISTA vs SwiftOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Qwen3.8-Flash-Next IQ3_S versus Swift Qwen3.8-Flash-Next IQ3_XXS
ISTA Qwen3.8-Flash-Next IQ3_S versus Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: ISTA; model B: Swift
Scope
Derived from primary HH2 captures; not another run
-3.680556 composite pointsPaired accuracy endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.493164 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-19.388941 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
13.250017 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
ISTA vs OrcaRouterOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Qwen3.8-Flash-Next IQ3_S versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
ISTA Qwen3.8-Flash-Next IQ3_S versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: ISTA; model B: OrcaRouter
Scope
Derived from primary HH2 captures; not another run
5.625 composite pointsPaired accuracy endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.371094 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-13.125 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
21.888906 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
ISTA vs Flash-Next UnslothOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Qwen3.8-Flash-Next IQ3_S versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
ISTA Qwen3.8-Flash-Next IQ3_S versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: ISTA; model B: Flash-Next Unsloth
Scope
Derived from primary HH2 captures; not another run
-0.486111 composite pointsPaired accuracy endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.881348 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-19.402795 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
17.583351 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
Swift vs OrcaRouterOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
Swift Qwen3.8-Flash-Next IQ3_XXS versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Swift; model B: OrcaRouter
Scope
Derived from primary HH2 captures; not another run
9.305556 composite pointsPaired accuracy endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.043945 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
0.878906 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-3.319462 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
21.694479 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
Swift vs Flash-Next UnslothOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Swift Qwen3.8-Flash-Next IQ3_XXS versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Swift; model B: Flash-Next Unsloth
Scope
Derived from primary HH2 captures; not another run
3.194444 composite pointsPaired accuracy endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.585938 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired accuracy endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired accuracy endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-13.027795 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
18.986128 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
OrcaRouter vs Flash-Next UnslothOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for OrcaRouter Qwen3.8-Flash-Next IQ3_M versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
OrcaRouter Qwen3.8-Flash-Next IQ3_M versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: OrcaRouter; model B: Flash-Next Unsloth
Scope
Derived from primary HH2 captures; not another run
Not recordedSource date above is not a capture date
Application / version
Strata native · version not recorded
Runtime / extension
Strata 0.1.40
Benchmark
HH2 paired statistics
Context capacity
262,144 tokens loaded
Thinking
Off
MTP
Per model and cohort; see recorded settings
K / V cache
Strata int8 or Prism Q8 K/V; see each result
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
10 results
Not applicable to statistical summaries · — = not available
Flash-Next HH2 · paired retention uncertainty. Scores are out of 100; No pooled quality score; signed model A minus model B effect with multiplicity-aware uncertainty. Time basis: Summed request minutes.
Model & configuration
Score / 100
Generation tok/s
Summed request minutes
Memory see basis
Evidence
Coder vs ISTAOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Coder Qwen3.8-Flash-Next IQ1_M versus ISTA Qwen3.8-Flash-Next IQ3_S
ISTA Coder Qwen3.8-Flash-Next IQ1_M versus ISTA Qwen3.8-Flash-Next IQ3_S
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Coder; model B: ISTA
Scope
Derived from primary HH2 captures; not another run
-1.791667 composite pointsPaired retention endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.708496 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired retention endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired retention endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired retention endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-8.041667 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
10.777778 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
Coder vs SwiftOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Coder Qwen3.8-Flash-Next IQ1_M versus Swift Qwen3.8-Flash-Next IQ3_XXS
ISTA Coder Qwen3.8-Flash-Next IQ1_M versus Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Coder; model B: Swift
Scope
Derived from primary HH2 captures; not another run
-5.652778 composite pointsPaired retention endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.271973 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired retention endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired retention endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired retention endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-18.138889 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
7.680573 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
Coder vs OrcaRouterOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Coder Qwen3.8-Flash-Next IQ1_M versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
ISTA Coder Qwen3.8-Flash-Next IQ1_M versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Coder; model B: OrcaRouter
Scope
Derived from primary HH2 captures; not another run
3.472222 composite pointsPaired retention endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.449707 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired retention endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired retention endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired retention endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-10.236146 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
14.777778 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
Coder vs Flash-Next UnslothOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Coder Qwen3.8-Flash-Next IQ1_M versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
ISTA Coder Qwen3.8-Flash-Next IQ1_M versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Coder; model B: Flash-Next Unsloth
Scope
Derived from primary HH2 captures; not another run
-2.180556 composite pointsPaired retention endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.717285 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired retention endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired retention endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired retention endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-17.375017 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
12.736128 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
ISTA vs SwiftOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Qwen3.8-Flash-Next IQ3_S versus Swift Qwen3.8-Flash-Next IQ3_XXS
ISTA Qwen3.8-Flash-Next IQ3_S versus Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: ISTA; model B: Swift
Scope
Derived from primary HH2 captures; not another run
-3.861111 composite pointsPaired retention endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.480469 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired retention endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired retention endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired retention endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-19.694444 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
13.166667 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
ISTA vs OrcaRouterOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Qwen3.8-Flash-Next IQ3_S versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
ISTA Qwen3.8-Flash-Next IQ3_S versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: ISTA; model B: OrcaRouter
Scope
Derived from primary HH2 captures; not another run
5.263889 composite pointsPaired retention endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.405273 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired retention endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired retention endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired retention endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-13.402795 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
21.625052 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
ISTA vs Flash-Next UnslothOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Qwen3.8-Flash-Next IQ3_S versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
ISTA Qwen3.8-Flash-Next IQ3_S versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: ISTA; model B: Flash-Next Unsloth
Scope
Derived from primary HH2 captures; not another run
-0.388889 composite pointsPaired retention endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.908691 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired retention endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired retention endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired retention endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-19.305556 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
17.555608 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
Swift vs OrcaRouterOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
Swift Qwen3.8-Flash-Next IQ3_XXS versus OrcaRouter Qwen3.8-Flash-Next IQ3_M
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Swift; model B: OrcaRouter
Scope
Derived from primary HH2 captures; not another run
9.125 composite pointsPaired retention endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.053223 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired retention endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired retention endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired retention endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-3.458351 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
21.694444 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
Swift vs Flash-Next UnslothOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Swift Qwen3.8-Flash-Next IQ3_XXS versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: Swift; model B: Flash-Next Unsloth
Scope
Derived from primary HH2 captures; not another run
3.472222 composite pointsPaired retention endpoint; Holm correction across all 20 declared comparisons
raw sign flip p
0.495117 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
holm adjusted p
1 probabilityPaired retention endpoint; Holm correction across all 20 declared comparisons
haystacks
12 countPaired retention endpoint; Holm correction across all 20 declared comparisons
sign flip assignments
4,096 countPaired retention endpoint; Holm correction across all 20 declared comparisons
bootstrap replicates
100,000 countPaired retention endpoint; Holm correction across all 20 declared comparisons
simultaneous approximate interval lower
-12.833333 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
simultaneous approximate interval upper
19.138889 composite pointsCluster bootstrap simultaneous approximate interval; same 12 haystacks
All comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
—
—
—
—
Descriptive · qualification limitsAll comparisons HOLD. No statistically separated pair. Effects are model A minus model B.
OrcaRouter vs Flash-Next UnslothOff · int8 / int8Captured: Not recordedCapture, runtime & configuration for OrcaRouter Qwen3.8-Flash-Next IQ3_M versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
OrcaRouter Qwen3.8-Flash-Next IQ3_M versus Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Test captured
Not recorded
Date precision
Selected request start/finish bounds; includes gaps, not summed duration
Capture timezone
Not recorded
Application / version
Strata native · version not recorded
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Five Flash-Next packages only. 480 primary requests plus 40 qualification. All 20 pairwise comparisons remain HOLD; no separated winner. Four models have repeat failures. Earlier approximately 240K diagnostic NOT RUN. Context 262144, actual longest input approximately 224K. Coder uses selected complete recovery. Twelve paired haystacks; exact sign-flip and 100,000 seeded cluster bootstrap resamples.
Thinking
Off
K / V cache
int8 / int8
MTP draft tokens
0
Comparison
Model A: OrcaRouter; model B: Flash-Next Unsloth
Scope
Derived from primary HH2 captures; not another run
XBench original full300 / original per-answer rubric
Context capacity
32,768 tokens
Thinking
Off
MTP
Per model and cohort; see recorded settings
K / V cache
Strata int8 or Prism Q8 K/V; see each result
900 targets;756 usable grades;144 ungraded. 300 row IDs but169 unique texts. One uncalibrated gpt-oss-safeguard-20b MXFP4 judge, medium reasoning, context65536/output32768. No full-suite semantic score. Subset percentages do not estimate missing responses. Strict refusal and weighted success are different.
3 results
Aggregate native target decode tokens / native target decode seconds; excludes judge, prefill and loading · — = not available
Flash-Next XBench · provisional graded subsets. Scores are out of 100; No overall score; read class-specific provisional refusal and usefulness metrics. Time basis: Summed request minutes.
Recorded package identities; installed weights were not rehashed during this review. Coder predates October17 update. RAM observed October11, not contemporaneous telemetry for every test.
7 results
Not applicable; use per-suite native timings · — = not available
Flash-Next · tested package identities and runtime. Scores are out of 100; No benchmark score; package storage and representative expert allocation are different quantities. Time basis: Summed request minutes.
Model & configuration
Score / 100
Generation tok/s
Summed request minutes
Memory see basis
Evidence
CoderPer cohort · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Coder Qwen3.8-Flash-Next IQ1_M
ISTA Coder Qwen3.8-Flash-Next IQ1_M
Test captured
Not recorded
Date precision
Package/configuration metadata, not a separate inference cohort
Capture timezone
Not recorded
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Recorded package identities; installed weights were not rehashed during this review. Coder predates October17 update. RAM observed October11, not contemporaneous telemetry for every test.
Thinking
Per cohort
K / V cache
int8 / int8
MTP draft tokens
Up to 2
Identity scope
Recorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review.
32,768 tokensRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
quant bench gpu expert cache gi b
21.4 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
quant bench pinned expert ram gi b
2.02 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy context tokens
98,304 tokensRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy gpu expert cache gi b
20.45 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy pinned expert ram gi b
2.97 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
Recorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review. Target GGUF shards; excludes runtime packs, draft/MTP files and runtime memory.
—
—
—
—
Recorded configurationRecorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review. Target GGUF shards; excludes runtime packs, draft/MTP files and runtime memory.
SwiftPer cohort · int8 / int8Captured: Not recordedCapture, runtime & configuration for Swift Qwen3.8-Flash-Next IQ3_XXS
Swift Qwen3.8-Flash-Next IQ3_XXS
Test captured
Not recorded
Date precision
Package/configuration metadata, not a separate inference cohort
Capture timezone
Not recorded
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Recorded package identities; installed weights were not rehashed during this review. Coder predates October17 update. RAM observed October11, not contemporaneous telemetry for every test.
Thinking
Per cohort
K / V cache
int8 / int8
MTP draft tokens
Up to 2
Identity scope
Recorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review.
32,768 tokensRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
quant bench gpu expert cache gi b
21.68 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
quant bench pinned expert ram gi b
18.28 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy context tokens
98,304 tokensRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy gpu expert cache gi b
20.72 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy pinned expert ram gi b
19.24 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
Recorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review. Target GGUF shards; excludes runtime packs, draft/MTP files and runtime memory.
—
—
—
—
Recorded configurationRecorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review. Target GGUF shards; excludes runtime packs, draft/MTP files and runtime memory.
ISTAPer cohort · int8 / int8Captured: Not recordedCapture, runtime & configuration for ISTA Qwen3.8-Flash-Next IQ3_S
ISTA Qwen3.8-Flash-Next IQ3_S
Test captured
Not recorded
Date precision
Package/configuration metadata, not a separate inference cohort
Capture timezone
Not recorded
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Recorded package identities; installed weights were not rehashed during this review. Coder predates October17 update. RAM observed October11, not contemporaneous telemetry for every test.
Thinking
Per cohort
K / V cache
int8 / int8
MTP draft tokens
Up to 2
Identity scope
Recorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review.
32,768 tokensRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
quant bench gpu expert cache gi b
21.34 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
quant bench pinned expert ram gi b
25.49 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy context tokens
98,304 tokensRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy gpu expert cache gi b
20.52 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy pinned expert ram gi b
26.32 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
Recorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review. Target GGUF shards; excludes runtime packs, draft/MTP files and runtime memory.
—
—
—
—
Recorded configurationRecorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review. Target GGUF shards; excludes runtime packs, draft/MTP files and runtime memory.
OrcaRouterPer cohort · int8 / int8Captured: Not recordedCapture, runtime & configuration for OrcaRouter Qwen3.8-Flash-Next IQ3_M
OrcaRouter Qwen3.8-Flash-Next IQ3_M
Test captured
Not recorded
Date precision
Package/configuration metadata, not a separate inference cohort
Capture timezone
Not recorded
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Recorded package identities; installed weights were not rehashed during this review. Coder predates October17 update. RAM observed October11, not contemporaneous telemetry for every test.
Thinking
Per cohort
K / V cache
int8 / int8
MTP draft tokens
Up to 2
Identity scope
Recorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review.
32,768 tokensRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
quant bench gpu expert cache gi b
21.97 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
quant bench pinned expert ram gi b
31.94 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy context tokens
98,304 tokensRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy gpu expert cache gi b
21.14 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy pinned expert ram gi b
32.76 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
Recorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review. Target GGUF shards; excludes runtime packs, draft/MTP files and runtime memory.
—
—
—
—
Recorded configurationRecorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review. Target GGUF shards; excludes runtime packs, draft/MTP files and runtime memory.
Flash-Next UnslothPer cohort · int8 / int8Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Unsloth Qwen3.8-Flash-Next UD-Q4_K_XL
Test captured
Not recorded
Date precision
Package/configuration metadata, not a separate inference cohort
Capture timezone
Not recorded
Application / version
Strata 0.1.40.1
Inference engine / build
Strata 0.1.40
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Recorded package identities; installed weights were not rehashed during this review. Coder predates October17 update. RAM observed October11, not contemporaneous telemetry for every test.
Thinking
Per cohort
K / V cache
int8 / int8
MTP draft tokens
0
Identity scope
Recorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review.
32,768 tokensRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
quant bench gpu expert cache gi b
21.23 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
quant bench pinned expert ram gi b
50.5 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy context tokens
98,304 tokensRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy gpu expert cache gi b
20.49 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
legacy pinned expert ram gi b
51.23 GiBRepresentative native startup expert allocations, rounded by the native logger; not total VRAM or whole-process RAM.
Recorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review. Target GGUF shards; excludes runtime packs, draft/MTP files and runtime memory.
—
—
—
—
Recorded configurationRecorded publisher inventory and native preparation/capture receipts; installed weights were not rehashed in this review. Target GGUF shards; excludes runtime packs, draft/MTP files and runtime memory.
27B UncensoredPer cohort · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for JonathanColetti Qwen3.8-27B Uncensored Q4_K_M
JonathanColetti Qwen3.8-27B Uncensored Q4_K_M
Test captured
Not recorded
Date precision
Package/configuration metadata, not a separate inference cohort
Capture timezone
Not recorded
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Recorded package identities; installed weights were not rehashed during this review. Coder predates October17 update. RAM observed October11, not contemporaneous telemetry for every test.
Thinking
Per cohort
K / V cache
Q8_0 / Q8_0
MTP draft tokens
0
Identity scope
Recorded local file bytes and SHA256 bound to native capture configuration; publisher revision not recorded here.
Recorded local file bytes and SHA256 bound to native capture configuration; publisher revision not recorded here. Target GGUF file, excluding runtime memory.
—
—
—
—
Recorded configurationRecorded local file bytes and SHA256 bound to native capture configuration; publisher revision not recorded here. Target GGUF file, excluding runtime memory.
27B StandardPer cohort · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B Standard UD-Q6_K
Unsloth Qwen3.8-27B Standard UD-Q6_K
Test captured
Not recorded
Date precision
Package/configuration metadata, not a separate inference cohort
Capture timezone
Not recorded
Application / version
Prism llama.cpp b10709
Inference engine / build
llama.cpp Prism fork b10709-9a9394a89
OS / environment
Intel Core i9-13900K; RTX 5090 32 GB; 128 GiB RAM
GPU driver version
Not recorded for this capture
CUDA runtime version
Not recorded for this capture
Comparison scope
Recorded package identities; installed weights were not rehashed during this review. Coder predates October17 update. RAM observed October11, not contemporaneous telemetry for every test.
Thinking
Per cohort
K / V cache
Q8_0 / Q8_0
MTP draft tokens
0
Identity scope
Recorded local file bytes and SHA256 bound to native capture configuration; publisher revision not recorded here.
Recorded local file bytes and SHA256 bound to native capture configuration; publisher revision not recorded here. Target GGUF file, excluding runtime memory.
—
—
—
—
Recorded configurationRecorded local file bytes and SHA256 bound to native capture configuration; publisher revision not recorded here. Target GGUF file, excluding runtime memory.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
QuantBench same 50 questions, seeds 5090–5094
Context capacity
65,536 tokens
Thinking
Off
MTP
Off
K / V cache
Q8_0 / Q8_0
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Same 50 questions, five seeds; no added independent questions.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · QuantBench Off, five passes. Scores are out of 100; QuantBench same 50 questions, seeds 5090–5094 · edition-specific score. Time basis: Not aggregated.
Model & configuration
Score / 100
Generation tok/s
Not aggregated
Memory see basis
Evidence
Qwen3.8-27B UD-Q6_KOff · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Same 50 questions, five seeds; no added independent questions.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Same 50 questions, five seeds; no added independent questions.
57.40
—
—
—
CompleteOfficial primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Same 50 questions, five seeds; no added independent questions.
Bolt 74B-A9BOff · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB mixed-precision GGUF
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Same 50 questions, five seeds; no added independent questions.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Same 50 questions, five seeds; no added independent questions.
44.40
—
—
—
CompleteOfficial primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Same 50 questions, five seeds; no added independent questions.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
Mirage neutral 300
Context capacity
65,536 tokens
Thinking
Off
MTP
Off
K / V cache
Q8_0 / Q8_0
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · Mirage neutral Off, format-confounded. Scores are out of 100; Mirage neutral 300 · edition-specific score. Time basis: Not aggregated.
Model & configuration
Score / 100
Generation tok/s
Not aggregated
Memory see basis
Evidence
Qwen3.8-27B UD-Q6_KOff · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
82.80
—
—
—
CompleteOfficial primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
Bolt 74B-A9BOff · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB mixed-precision GGUF
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
8.70
—
—
—
CompleteOfficial primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Bolt vs Qwen · Mirage pressure Off, format-confounded
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
Mirage pressure 300
Context capacity
65,536 tokens
Thinking
Off
MTP
Off
K / V cache
Q8_0 / Q8_0
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · Mirage pressure Off, format-confounded. Scores are out of 100; Mirage pressure 300 · edition-specific score. Time basis: Not aggregated.
Model & configuration
Score / 100
Generation tok/s
Not aggregated
Memory see basis
Evidence
Qwen3.8-27B UD-Q6_KOff · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
82.60
—
—
—
CompleteOfficial primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
Bolt 74B-A9BOff · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB mixed-precision GGUF
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
54.40
—
—
—
CompleteOfficial primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. FORMAT-CONFOUNDED: template echo failure; not a measured hallucination rate.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
Evidence-to-Action evidence18
Context capacity
65,536 tokens
Thinking
xHigh
MTP
Off
K / V cache
Q8_0 / Q8_0
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Official 16,384-token cap. Bolt capped on 3/18 requests; do not substitute the 32K diagnostic.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · Evidence-to-Action xHigh, official 16K cap. Scores are out of 100; Evidence-to-Action evidence18 · edition-specific score. Time basis: Not aggregated.
Model & configuration
Score / 100
Generation tok/s
Not aggregated
Memory see basis
Evidence
Qwen3.8-27B UD-Q6_KxHigh · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Official 16,384-token cap. Bolt capped on 3/18 requests; do not substitute the 32K diagnostic.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Official 16,384-token cap. Bolt capped on 3/18 requests; do not substitute the 32K diagnostic.
98.50
—
—
—
CompleteOfficial primary comparison. tied. Each test retains its own verdict; no composite ranking. Official 16,384-token cap. Bolt capped on 3/18 requests; do not substitute the 32K diagnostic.
Bolt 74B-A9BxHigh · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB mixed-precision GGUF
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Official 16,384-token cap. Bolt capped on 3/18 requests; do not substitute the 32K diagnostic.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Official 16,384-token cap. Bolt capped on 3/18 requests; do not substitute the 32K diagnostic.
83.00
—
—
—
CompleteOfficial primary comparison. tied. Each test retains its own verdict; no composite ranking. Official 16,384-token cap. Bolt capped on 3/18 requests; do not substitute the 32K diagnostic.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
QuantBench same 50 questions, seed 5090
Context capacity
65,536 tokens
Thinking
xHigh
MTP
Off
K / V cache
Q8_0 / Q8_0
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. One pass only; additional thinking-on passes not run. On/Off differences descriptive because Qwen samplers differ.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · QuantBench xHigh, one pass. Scores are out of 100; QuantBench same 50 questions, seed 5090 · edition-specific score. Time basis: Not aggregated.
Model & configuration
Score / 100
Generation tok/s
Not aggregated
Memory see basis
Evidence
Qwen3.8-27B UD-Q6_KxHigh · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. One pass only; additional thinking-on passes not run. On/Off differences descriptive because Qwen samplers differ.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. One pass only; additional thinking-on passes not run. On/Off differences descriptive because Qwen samplers differ.
80.60
—
—
—
CompleteOfficial primary comparison. tied. Each test retains its own verdict; no composite ranking. One pass only; additional thinking-on passes not run. On/Off differences descriptive because Qwen samplers differ.
Bolt 74B-A9BxHigh · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB mixed-precision GGUF
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. One pass only; additional thinking-on passes not run. On/Off differences descriptive because Qwen samplers differ.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. One pass only; additional thinking-on passes not run. On/Off differences descriptive because Qwen samplers differ.
71.60
—
—
—
CompleteOfficial primary comparison. tied. Each test retains its own verdict; no composite ranking. One pass only; additional thinking-on passes not run. On/Off differences descriptive because Qwen samplers differ.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
LiveCodeBench v6, first 70 of 175 in frozen order
Context capacity
65,536 tokens
Thinking
xHigh
MTP
Off
K / V cache
Q8_0 / Q8_0
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Outcome-blind time-based cutoff; nonrandom first 70. 32K output cap. Bolt 17 and Qwen 19 out of budget; wrong code 6 and 5.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · LiveCodeBench xHigh, first 70. Scores are out of 70; LiveCodeBench v6, first 70 of 175 in frozen order · paired solved count. Time basis: Not aggregated.
Model & configuration
Score / 70
Generation tok/s
Not aggregated
Memory see basis
Evidence
Bolt 74B-A9BxHigh · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB mixed-precision GGUF
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Outcome-blind time-based cutoff; nonrandom first 70. 32K output cap. Bolt 17 and Qwen 19 out of budget; wrong code 6 and 5.
All measurements and denominators (2)
out of budget
17 problemsFirst 70, still thinking at 32K cap
wrong code
6 problemsFirst 70
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Outcome-blind time-based cutoff; nonrandom first 70. 32K output cap. Bolt 17 and Qwen 19 out of budget; wrong code 6 and 5.
47.00
—
—
—
CompleteOfficial primary comparison. tied. Each test retains its own verdict; no composite ranking. Outcome-blind time-based cutoff; nonrandom first 70. 32K output cap. Bolt 17 and Qwen 19 out of budget; wrong code 6 and 5.
Qwen3.8-27B UD-Q6_KxHigh · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Outcome-blind time-based cutoff; nonrandom first 70. 32K output cap. Bolt 17 and Qwen 19 out of budget; wrong code 6 and 5.
All measurements and denominators (2)
out of budget
19 problemsFirst 70, still thinking at 32K cap
wrong code
5 problemsFirst 70
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Outcome-blind time-based cutoff; nonrandom first 70. 32K output cap. Bolt 17 and Qwen 19 out of budget; wrong code 6 and 5.
46.00
—
—
—
CompleteOfficial primary comparison. tied. Each test retains its own verdict; no composite ranking. Outcome-blind time-based cutoff; nonrandom first 70. 32K output cap. Bolt 17 and Qwen 19 out of budget; wrong code 6 and 5.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
HH2-1.0 frozen profile
Context capacity
32K context tier
Thinking
Off; deterministic HH2 profile
MTP
Off
K / V cache
Q8_0 / Q8_0
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · HH-2.0 32K, primary six haystacks. Scores are out of 100; HH2-1.0 frozen profile · edition-specific score. Time basis: Not aggregated.
Model & configuration
Score / 100
Generation tok/s
Not aggregated
Memory see basis
Evidence
Qwen3.8-27B UD-Q6_KOff; deterministic HH2 profile · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
97.40
—
—
—
CompleteOfficial primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
Official primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
93.70
—
—
—
CompleteOfficial primary comparison. Qwen ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
HH2-1.0 frozen profile
Context capacity
64K context tier
Thinking
Off; deterministic HH2 profile
MTP
Off
K / V cache
Q8_0 / Q8_0
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · HH-2.0 64K, primary six haystacks. Scores are out of 100; HH2-1.0 frozen profile · edition-specific score. Time basis: Not aggregated.
Model & configuration
Score / 100
Generation tok/s
Not aggregated
Memory see basis
Evidence
Qwen3.8-27B UD-Q6_KOff; deterministic HH2 profile · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
96.00
—
—
—
CompleteOfficial primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
92.70
—
—
—
CompleteOfficial primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
HH2-1.0 frozen profile
Context capacity
128K context tier
Thinking
Off; deterministic HH2 profile
MTP
Off
K / V cache
Q8_0 / Q8_0
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · HH-2.0 128K, primary six haystacks. Scores are out of 100; HH2-1.0 frozen profile · edition-specific score. Time basis: Not aggregated.
Model & configuration
Score / 100
Generation tok/s
Not aggregated
Memory see basis
Evidence
Qwen3.8-27B UD-Q6_KOff; deterministic HH2 profile · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
89.30
—
—
—
CompleteOfficial primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
Official primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
84.90
—
—
—
CompleteOfficial primary comparison. tied. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
HH2-1.0 frozen profile
Context capacity
256K context tier
Thinking
Off; deterministic HH2 profile
MTP
Off
K / V cache
Q8_0 / Q8_0
Official primary comparison. Bolt ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc. Qwen micro-batch 256; native micro-batch 512 failed the 1 GiB headroom rule.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · HH-2.0 256K, primary six haystacks. Scores are out of 100; HH2-1.0 frozen profile · edition-specific score. Time basis: Not aggregated.
Official primary comparison. Bolt ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc. Qwen micro-batch 256; native micro-batch 512 failed the 1 GiB headroom rule.
Official primary comparison. Bolt ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc. Qwen micro-batch 256; native micro-batch 512 failed the 1 GiB headroom rule.
79.40
—
—
—
CompleteOfficial primary comparison. Bolt ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc. Qwen micro-batch 256; native micro-batch 512 failed the 1 GiB headroom rule.
Qwen3.8-27B UD-Q6_KOff; deterministic HH2 profile · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Official primary comparison. Bolt ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc. Qwen micro-batch 256; native micro-batch 512 failed the 1 GiB headroom rule.
Official primary comparison. Bolt ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc. Qwen micro-batch 256; native micro-batch 512 failed the 1 GiB headroom rule.
65.20
—
—
—
CompleteOfficial primary comparison. Bolt ahead. Each test retains its own verdict; no composite ranking. Six haystacks per tier; Classic cold, paired Reasoning reuses cached haystack prefix. Primary result; the later eight-haystack extension is post hoc. Qwen micro-batch 256; native micro-batch 512 failed the 1 GiB headroom rule.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
Evidence-to-Action evidence18
Context capacity
65,536 tokens
Thinking
xHigh
MTP
Off
K / V cache
Q8_0 / Q8_0
DIAGNOSTIC SENSITIVITY: official cap is 16,384. Decisions tied; citations Qwen ahead. Only the output cap changed to 32,768.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · E2A xHigh, 32K-cap DIAGNOSTIC. Scores are out of 100; Evidence-to-Action evidence18 · edition-specific score. Time basis: Not aggregated.
Model & configuration
Score / 100
Generation tok/s
Not aggregated
Memory see basis
Evidence
Bolt 74B-A9BxHigh · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB mixed-precision GGUF
Not recordedSource date above is not a capture date
Application / version
llama-server b11669
Runtime / extension
llama.cpp b11669
Benchmark
Mirage neutral lower-bound normalization
Context capacity
65,536 tokens
Thinking
Off
MTP
Off
K / V cache
Q8_0 / Q8_0
Post hoc lower-bound DIAGNOSTIC; official score unchanged. Only echoed verdict lines normalized. Qwen ahead overall; fake resistance tied.
2 results
Not measured in this quality summary · — = not available
Bolt vs Qwen · Mirage neutral, post hoc lower-bound DIAGNOSTIC. Scores are out of 100; Mirage neutral lower-bound normalization · edition-specific score. Time basis: Not aggregated.
Model & configuration
Score / 100
Generation tok/s
Not aggregated
Memory see basis
Evidence
Qwen3.8-27B UD-Q6_KOff · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Serving performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
2 results
Mean decode tokens/s, SPEED-Bench qualitative coding · — = not available
Bolt vs Qwen · Native decoding. Scores are out of 100; SPEED-Bench qualitative coding, 80 conversations, 89 turns · edition-specific score. Time basis: Not aggregated.
Serving performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
All measurements and denominators (2)
conversation time
14.6 secondsMean conversation wall time; not task correctness
tokens per conversation
2,221 tokensMean
Serving performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
—
154.40
—
—
CompleteServing performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
Qwen3.8-27B UD-Q6_KProfile-specific; read speed conditions · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Serving performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
All measurements and denominators (2)
conversation time
73.9 secondsMean conversation wall time; not task correctness
tokens per conversation
4,608 tokensMean
Serving performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
—
62.70
—
—
CompleteServing performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Serving performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
2 results
Mean decode tokens/s, SPEED-Bench qualitative coding · — = not available
Bolt vs Qwen · Both accelerated. Scores are out of 100; SPEED-Bench qualitative coding, 80 conversations, 89 turns · edition-specific score. Time basis: Not aggregated.
Serving performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
All measurements and denominators (1)
conversation time
7.7 secondsMean conversation wall time; not task correctness
Serving performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
—
347.90
—
—
CompleteServing performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
Qwen3.8-27B UD-Q6_KProfile-specific; read speed conditions · Q8_0 / Q8_0Captured: Not recordedCapture, runtime & configuration for Unsloth Qwen3.8-27B UD-Q6_K
Unsloth Qwen3.8-27B UD-Q6_K
Test captured
Not recorded
Date precision
Per-request capture dates are not included in this whitepaper summary.
Serving performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
All measurements and denominators (1)
conversation time
36 secondsMean conversation wall time; not task correctness
Serving performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
—
127.10
—
—
CompleteServing performance only, content not graded; cap-hit turns included. Compare native with native or both accelerated. Not time to a correct answer.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Bolt vs Qwen / Reviewed: 28 Oct 2026; results documented 19 Oct 2026
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
8 results
Not measured in this profile · — = not available
KV cache · 4K · Asset registry. Scores are out of 100; Exact-match accuracy percent; fixed parser, one task and context bucket. Time basis: Not reported for this profile.
100 questionsOfficial accepted questions; null when unavailable
correct
100 answersSaved deterministic grade
wilson95 low
96.30065 percentPer-cell Wilson 95% interval
wilson95 high
100 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
99 answersSaved deterministic grade
wilson95 low
94.55138 percentPer-cell Wilson 95% interval
wilson95 high
99.823257 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
99.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
99 answersSaved deterministic grade
wilson95 low
94.55138 percentPer-cell Wilson 95% interval
wilson95 high
99.823257 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
99.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
99 answersSaved deterministic grade
wilson95 low
94.55138 percentPer-cell Wilson 95% interval
wilson95 high
99.823257 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
99.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
96 answersSaved deterministic grade
wilson95 low
90.162929 percentPer-cell Wilson 95% interval
wilson95 high
98.43367 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
4 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
96.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
96 answersSaved deterministic grade
wilson95 low
90.162929 percentPer-cell Wilson 95% interval
wilson95 high
98.43367 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
4 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
96.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
96 answersSaved deterministic grade
wilson95 low
90.162929 percentPer-cell Wilson 95% interval
wilson95 high
98.43367 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
4 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
96.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
94 answersSaved deterministic grade
wilson95 low
87.523185 percentPer-cell Wilson 95% interval
wilson95 high
97.221388 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
1 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
5 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
94.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
KV cache / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 4K · Config tracking
Exact-match accuracy on config retrieval; 100 questions per available model/cache task cell.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
8 results
Not measured in this profile · — = not available
KV cache · 4K · Config tracking. Scores are out of 100; Exact-match accuracy percent; fixed parser, one task and context bucket. Time basis: Not reported for this profile.
100 questionsOfficial accepted questions; null when unavailable
correct
94 answersSaved deterministic grade
wilson95 low
87.523185 percentPer-cell Wilson 95% interval
wilson95 high
97.221388 percentPer-cell Wilson 95% interval
error format
2 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
4 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
94.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
91 answersSaved deterministic grade
wilson95 low
83.773787 percentPer-cell Wilson 95% interval
wilson95 high
95.192746 percentPer-cell Wilson 95% interval
error format
7 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
2 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
91.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
86 answersSaved deterministic grade
wilson95 low
77.862812 percentPer-cell Wilson 95% interval
wilson95 high
91.473656 percentPer-cell Wilson 95% interval
error format
12 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
2 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
86.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
86 answersSaved deterministic grade
wilson95 low
77.862812 percentPer-cell Wilson 95% interval
wilson95 high
91.473656 percentPer-cell Wilson 95% interval
error format
11 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
3 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
86.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
79 answersSaved deterministic grade
wilson95 low
70.020031 percentPer-cell Wilson 95% interval
wilson95 high
85.834346 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
20 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
79.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
79 answersSaved deterministic grade
wilson95 low
70.020031 percentPer-cell Wilson 95% interval
wilson95 high
85.834346 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
20 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
79.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
78 answersSaved deterministic grade
wilson95 low
68.929646 percentPer-cell Wilson 95% interval
wilson95 high
84.998718 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
22 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
78.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
72 answersSaved deterministic grade
wilson95 low
62.511971 percentPer-cell Wilson 95% interval
wilson95 high
79.860315 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
26 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
2 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
72.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
KV cache / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 16K · Asset registry
Exact-match accuracy on asset retrieval; 100 questions per available model/cache task cell.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
8 results
Not measured in this profile · — = not available
KV cache · 16K · Asset registry. Scores are out of 100; Exact-match accuracy percent; fixed parser, one task and context bucket. Time basis: Not reported for this profile.
100 questionsOfficial accepted questions; null when unavailable
correct
100 answersSaved deterministic grade
wilson95 low
96.30065 percentPer-cell Wilson 95% interval
wilson95 high
100 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
100 answersSaved deterministic grade
wilson95 low
96.30065 percentPer-cell Wilson 95% interval
wilson95 high
100 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
100 answersSaved deterministic grade
wilson95 low
96.30065 percentPer-cell Wilson 95% interval
wilson95 high
100 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
100 answersSaved deterministic grade
wilson95 low
96.30065 percentPer-cell Wilson 95% interval
wilson95 high
100 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
91 answersSaved deterministic grade
wilson95 low
83.773787 percentPer-cell Wilson 95% interval
wilson95 high
95.192746 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
1 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
8 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
91.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
89 answersSaved deterministic grade
wilson95 low
81.368703 percentPer-cell Wilson 95% interval
wilson95 high
93.745804 percentPer-cell Wilson 95% interval
error format
1 answersCanonical saved error class; no regrading
error near collision
1 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
9 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
89.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
87 answersSaved deterministic grade
wilson95 low
79.019649 percentPer-cell Wilson 95% interval
wilson95 high
92.242833 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
2 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
11 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
87.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
87 answersSaved deterministic grade
wilson95 low
79.019649 percentPer-cell Wilson 95% interval
wilson95 high
92.242833 percentPer-cell Wilson 95% interval
error format
2 answersCanonical saved error class; no regrading
error near collision
1 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
10 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
87.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
KV cache / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 16K · Config tracking
Exact-match accuracy on config retrieval; 100 questions per available model/cache task cell.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
8 results
Not measured in this profile · — = not available
KV cache · 16K · Config tracking. Scores are out of 100; Exact-match accuracy percent; fixed parser, one task and context bucket. Time basis: Not reported for this profile.
100 questionsOfficial accepted questions; null when unavailable
correct
85 answersSaved deterministic grade
wilson95 low
76.71644 percentPer-cell Wilson 95% interval
wilson95 high
90.694015 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
13 answersCanonical saved error class; no regrading
error neighbor
2 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
85.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
84 answersSaved deterministic grade
wilson95 low
75.579731 percentPer-cell Wilson 95% interval
wilson95 high
89.904712 percentPer-cell Wilson 95% interval
error format
2 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
13 answersCanonical saved error class; no regrading
error neighbor
1 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
84.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
83 answersSaved deterministic grade
wilson95 low
74.451995 percentPer-cell Wilson 95% interval
wilson95 high
89.106434 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
16 answersCanonical saved error class; no regrading
error neighbor
1 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
83.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
83 answersSaved deterministic grade
wilson95 low
74.451995 percentPer-cell Wilson 95% interval
wilson95 high
89.106434 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
16 answersCanonical saved error class; no regrading
error neighbor
1 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
83.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
36 answersSaved deterministic grade
wilson95 low
27.27122 percentPer-cell Wilson 95% interval
wilson95 high
45.764598 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
61 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
3 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
36.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
35 answersSaved deterministic grade
wilson95 low
26.364248 percentPer-cell Wilson 95% interval
wilson95 high
44.745557 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
62 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
3 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
35.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
35 answersSaved deterministic grade
wilson95 low
26.364248 percentPer-cell Wilson 95% interval
wilson95 high
44.745557 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
63 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
2 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
35.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
34 answersSaved deterministic grade
wilson95 low
25.461521 percentPer-cell Wilson 95% interval
wilson95 high
43.722271 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
62 answersCanonical saved error class; no regrading
error neighbor
1 answersCanonical saved error class; no regrading
error other in doc
3 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
34.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
KV cache / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 32K · Asset registry
Exact-match accuracy on asset retrieval; 100 questions per available model/cache task cell.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
8 results
Not measured in this profile · — = not available
KV cache · 32K · Asset registry. Scores are out of 100; Exact-match accuracy percent; fixed parser, one task and context bucket. Time basis: Not reported for this profile.
100 questionsOfficial accepted questions; null when unavailable
correct
98 answersSaved deterministic grade
wilson95 low
92.998821 percentPer-cell Wilson 95% interval
wilson95 high
99.449803 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
2 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
98.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
98 answersSaved deterministic grade
wilson95 low
92.998821 percentPer-cell Wilson 95% interval
wilson95 high
99.449803 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
2 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
98.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
97 answersSaved deterministic grade
wilson95 low
91.548064 percentPer-cell Wilson 95% interval
wilson95 high
98.974548 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
3 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
97.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
97 answersSaved deterministic grade
wilson95 low
91.548064 percentPer-cell Wilson 95% interval
wilson95 high
98.974548 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
3 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
97.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
73 answersSaved deterministic grade
wilson95 low
63.567883 percentPer-cell Wilson 95% interval
wilson95 high
80.730416 percentPer-cell Wilson 95% interval
error format
8 answersCanonical saved error class; no regrading
error near collision
5 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
14 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
73.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
62 answersSaved deterministic grade
wilson95 low
52.209755 percentPer-cell Wilson 95% interval
wilson95 high
70.902401 percentPer-cell Wilson 95% interval
error format
32 answersCanonical saved error class; no regrading
error near collision
1 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
5 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
62.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
62 answersSaved deterministic grade
wilson95 low
52.209755 percentPer-cell Wilson 95% interval
wilson95 high
70.902401 percentPer-cell Wilson 95% interval
error format
30 answersCanonical saved error class; no regrading
error near collision
1 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
7 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
62.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
60 answersSaved deterministic grade
wilson95 low
50.200259 percentPer-cell Wilson 95% interval
wilson95 high
69.059871 percentPer-cell Wilson 95% interval
error format
33 answersCanonical saved error class; no regrading
error near collision
1 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
6 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
60.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
KV cache / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 32K · Config tracking
Exact-match accuracy on config retrieval; 100 questions per available model/cache task cell.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
8 results
Not measured in this profile · — = not available
KV cache · 32K · Config tracking. Scores are out of 100; Exact-match accuracy percent; fixed parser, one task and context bucket. Time basis: Not reported for this profile.
100 questionsOfficial accepted questions; null when unavailable
correct
75 answersSaved deterministic grade
wilson95 low
65.695536 percentPer-cell Wilson 95% interval
wilson95 high
82.454789 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
22 answersCanonical saved error class; no regrading
error neighbor
3 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
75.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
75 answersSaved deterministic grade
wilson95 low
65.695536 percentPer-cell Wilson 95% interval
wilson95 high
82.454789 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
20 answersCanonical saved error class; no regrading
error neighbor
5 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
75.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
74 answersSaved deterministic grade
wilson95 low
64.629011 percentPer-cell Wilson 95% interval
wilson95 high
81.595302 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
21 answersCanonical saved error class; no regrading
error neighbor
5 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
74.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
74 answersSaved deterministic grade
wilson95 low
64.629011 percentPer-cell Wilson 95% interval
wilson95 high
81.595302 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
21 answersCanonical saved error class; no regrading
error neighbor
5 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
74.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
35 answersSaved deterministic grade
wilson95 low
26.364248 percentPer-cell Wilson 95% interval
wilson95 high
44.745557 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
63 answersCanonical saved error class; no regrading
error neighbor
1 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
35.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
34 answersSaved deterministic grade
wilson95 low
25.461521 percentPer-cell Wilson 95% interval
wilson95 high
43.722271 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
63 answersCanonical saved error class; no regrading
error neighbor
2 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
34.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
34 answersSaved deterministic grade
wilson95 low
25.461521 percentPer-cell Wilson 95% interval
wilson95 high
43.722271 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
63 answersCanonical saved error class; no regrading
error neighbor
2 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
34.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
33 answersSaved deterministic grade
wilson95 low
24.563123 percentPer-cell Wilson 95% interval
wilson95 high
42.694656 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
64 answersCanonical saved error class; no regrading
error neighbor
2 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
33.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
KV cache / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 64K · Asset registry
Exact-match accuracy on asset retrieval; 100 questions per available model/cache task cell.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
8 results
Not measured in this profile · — = not available
KV cache · 64K · Asset registry. Scores are out of 100; Exact-match accuracy percent; fixed parser, one task and context bucket. Time basis: Not reported for this profile.
100 questionsOfficial accepted questions; null when unavailable
correct
100 answersSaved deterministic grade
wilson95 low
96.30065 percentPer-cell Wilson 95% interval
wilson95 high
100 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
99 answersSaved deterministic grade
wilson95 low
94.55138 percentPer-cell Wilson 95% interval
wilson95 high
99.823257 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
99.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
99 answersSaved deterministic grade
wilson95 low
94.55138 percentPer-cell Wilson 95% interval
wilson95 high
99.823257 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
99.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
99 answersSaved deterministic grade
wilson95 low
94.55138 percentPer-cell Wilson 95% interval
wilson95 high
99.823257 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
1 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
99.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
63 answersSaved deterministic grade
wilson95 low
53.22053 percentPer-cell Wilson 95% interval
wilson95 high
71.817639 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
13 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
22 answersCanonical saved error class; no regrading
error hallucinated
2 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
63.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
61 answersSaved deterministic grade
wilson95 low
51.203017 percentPer-cell Wilson 95% interval
wilson95 high
69.983126 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
15 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
23 answersCanonical saved error class; no regrading
error hallucinated
1 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
61.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
60 answersSaved deterministic grade
wilson95 low
50.200259 percentPer-cell Wilson 95% interval
wilson95 high
69.059871 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
16 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
23 answersCanonical saved error class; no regrading
error hallucinated
1 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
60.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
58 answersSaved deterministic grade
wilson95 low
48.206487 percentPer-cell Wilson 95% interval
wilson95 high
67.201617 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
16 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
24 answersCanonical saved error class; no regrading
error hallucinated
2 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
58.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
KV cache / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 64K · Config tracking
Exact-match accuracy on config retrieval; 100 questions per available model/cache task cell.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
8 results
Not measured in this profile · — = not available
KV cache · 64K · Config tracking. Scores are out of 100; Exact-match accuracy percent; fixed parser, one task and context bucket. Time basis: Not reported for this profile.
100 questionsOfficial accepted questions; null when unavailable
correct
73 answersSaved deterministic grade
wilson95 low
63.567883 percentPer-cell Wilson 95% interval
wilson95 high
80.730416 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
22 answersCanonical saved error class; no regrading
error neighbor
5 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
73.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
73 answersSaved deterministic grade
wilson95 low
63.567883 percentPer-cell Wilson 95% interval
wilson95 high
80.730416 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
22 answersCanonical saved error class; no regrading
error neighbor
5 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
73.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
71 answersSaved deterministic grade
wilson95 low
61.461109 percentPer-cell Wilson 95% interval
wilson95 high
78.985164 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
22 answersCanonical saved error class; no regrading
error neighbor
7 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
71.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
67 answersSaved deterministic grade
wilson95 low
57.305344 percentPer-cell Wilson 95% interval
wilson95 high
75.436877 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
28 answersCanonical saved error class; no regrading
error neighbor
5 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
67.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
36 answersSaved deterministic grade
wilson95 low
27.27122 percentPer-cell Wilson 95% interval
wilson95 high
45.764598 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
60 answersCanonical saved error class; no regrading
error neighbor
2 answersCanonical saved error class; no regrading
error other in doc
2 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
36.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
32 answersSaved deterministic grade
wilson95 low
23.669147 percentPer-cell Wilson 95% interval
wilson95 high
41.662619 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
63 answersCanonical saved error class; no regrading
error neighbor
2 answersCanonical saved error class; no regrading
error other in doc
3 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
32.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
29 answersSaved deterministic grade
wilson95 low
21.014836 percentPer-cell Wilson 95% interval
wilson95 high
38.538891 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
66 answersCanonical saved error class; no regrading
error neighbor
2 answersCanonical saved error class; no regrading
error other in doc
3 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
29.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
29 answersSaved deterministic grade
wilson95 low
21.014836 percentPer-cell Wilson 95% interval
wilson95 high
38.538891 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
66 answersCanonical saved error class; no regrading
error neighbor
2 answersCanonical saved error class; no regrading
error other in doc
3 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
29.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
KV cache / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 128K · Asset registry
Exact-match accuracy on asset retrieval; 100 questions per available model/cache task cell.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
8 results
Not measured in this profile · — = not available
KV cache · 128K · Asset registry. Scores are out of 100; Exact-match accuracy percent; fixed parser, one task and context bucket. Time basis: Not reported for this profile.
100 questionsOfficial accepted questions; null when unavailable
correct
98 answersSaved deterministic grade
wilson95 low
92.998821 percentPer-cell Wilson 95% interval
wilson95 high
99.449803 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
2 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
98.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
97 answersSaved deterministic grade
wilson95 low
91.548064 percentPer-cell Wilson 95% interval
wilson95 high
98.974548 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
3 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
97.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
97 answersSaved deterministic grade
wilson95 low
91.548064 percentPer-cell Wilson 95% interval
wilson95 high
98.974548 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
3 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
97.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
95 answersSaved deterministic grade
wilson95 low
88.824953 percentPer-cell Wilson 95% interval
wilson95 high
97.845632 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
5 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
95.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
51 answersSaved deterministic grade
wilson95 low
41.348012 percentPer-cell Wilson 95% interval
wilson95 high
60.578001 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
10 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
37 answersCanonical saved error class; no regrading
error hallucinated
2 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
51.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
51 answersSaved deterministic grade
wilson95 low
41.348012 percentPer-cell Wilson 95% interval
wilson95 high
60.578001 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
9 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
36 answersCanonical saved error class; no regrading
error hallucinated
4 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
51.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
49 answersSaved deterministic grade
wilson95 low
39.421999 percentPer-cell Wilson 95% interval
wilson95 high
58.651988 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
9 answersCanonical saved error class; no regrading
error stale
0 answersCanonical saved error class; no regrading
error neighbor
0 answersCanonical saved error class; no regrading
error other in doc
40 answersCanonical saved error class; no regrading
error hallucinated
2 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
49.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Not recordedCapture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Not recorded
Date precision
No accepted captures for this configuration.
Capture timezone
Not recorded
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
No accepted allocation
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
8 results
Not measured in this profile · — = not available
KV cache · 128K · Config tracking. Scores are out of 100; Exact-match accuracy percent; fixed parser, one task and context bucket. Time basis: Not reported for this profile.
100 questionsOfficial accepted questions; null when unavailable
correct
68 answersSaved deterministic grade
wilson95 low
58.337381 percentPer-cell Wilson 95% interval
wilson95 high
76.330853 percentPer-cell Wilson 95% interval
error format
2 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
27 answersCanonical saved error class; no regrading
error neighbor
3 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
68.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
67 answersSaved deterministic grade
wilson95 low
57.305344 percentPer-cell Wilson 95% interval
wilson95 high
75.436877 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
30 answersCanonical saved error class; no regrading
error neighbor
3 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
67.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
66 answersSaved deterministic grade
wilson95 low
56.277729 percentPer-cell Wilson 95% interval
wilson95 high
74.538479 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
31 answersCanonical saved error class; no regrading
error neighbor
3 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
66.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
100 questionsOfficial accepted questions; null when unavailable
correct
64 answersSaved deterministic grade
wilson95 low
54.235402 percentPer-cell Wilson 95% interval
wilson95 high
72.72878 percentPer-cell Wilson 95% interval
error format
3 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
30 answersCanonical saved error class; no regrading
error neighbor
3 answersCanonical saved error class; no regrading
error other in doc
0 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
64.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
31 answersSaved deterministic grade
wilson95 low
22.779697 percentPer-cell Wilson 95% interval
wilson95 high
40.626056 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
58 answersCanonical saved error class; no regrading
error neighbor
7 answersCanonical saved error class; no regrading
error other in doc
4 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
31.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
30 answersSaved deterministic grade
wilson95 low
21.894885 percentPer-cell Wilson 95% interval
wilson95 high
39.584855 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
57 answersCanonical saved error class; no regrading
error neighbor
8 answersCanonical saved error class; no regrading
error other in doc
5 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
30.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
100 questionsOfficial accepted questions; null when unavailable
correct
28 answersSaved deterministic grade
wilson95 low
20.139685 percentPer-cell Wilson 95% interval
wilson95 high
37.488029 percentPer-cell Wilson 95% interval
error format
0 answersCanonical saved error class; no regrading
error near collision
0 answersCanonical saved error class; no regrading
error stale
58 answersCanonical saved error class; no regrading
error neighbor
9 answersCanonical saved error class; no regrading
error other in doc
5 answersCanonical saved error class; no regrading
error hallucinated
0 answersCanonical saved error class; no regrading
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
28.00
—
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Not recordedCapture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Not recorded
Date precision
No accepted captures for this configuration.
Capture timezone
Not recorded
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
No accepted allocation
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects. Answers capped at 48 tokens. Cold samples are documents, not all scored questions.
8 results
Median per-request generation tok/s from accepted short answers; not sustained decode · — = not available
KV cache · 4K · short-answer speed. Scores are out of 100; No primary score; named metrics retain their own units. Time basis: Not reported for this profile.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,778.265604 tok/sAccepted launches; task-pooled sample population
cold samples
40 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
131.575704 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
129.499883 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
131.58
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,773.357935 tok/sAccepted launches; task-pooled sample population
cold samples
40 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
127.483957 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
125.638315 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
127.48
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,766.466874 tok/sAccepted launches; task-pooled sample population
cold samples
40 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
126.934764 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
125.824891 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
126.93
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,770.11262 tok/sAccepted launches; task-pooled sample population
cold samples
40 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
127.157712 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
124.991445 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
127.16
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
4,190.609055 tok/sAccepted launches; task-pooled sample population
cold samples
40 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
72.261934 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
72.371696 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
72.26
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
4,062.188445 tok/sAccepted launches; task-pooled sample population
cold samples
40 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
69.55466 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
69.629797 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
69.55
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
4,064.265974 tok/sAccepted launches; task-pooled sample population
cold samples
40 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
69.612102 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
69.693446 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
69.61
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
4,067.280568 tok/sAccepted launches; task-pooled sample population
cold samples
40 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
65.669227 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
65.899177 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
65.67
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Performance / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 16K · short-answer speed
Median cold document-processing and answer-generation rates; both tasks pooled.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects. Answers capped at 48 tokens. Cold samples are documents, not all scored questions.
8 results
Median per-request generation tok/s from accepted short answers; not sustained decode · — = not available
KV cache · 16K · short-answer speed. Scores are out of 100; No primary score; named metrics retain their own units. Time basis: Not reported for this profile.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,934.68524 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
128.2226 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
127.345042 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
128.22
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,925.561168 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
124.358788 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
123.705309 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
124.36
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,925.030214 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
122.237249 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
120.509148 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
122.24
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,919.327182 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
123.464413 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
121.954983 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
123.46
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
3,917.424357 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
66.443956 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
66.521944 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
66.44
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
3,833.280446 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
65.074211 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
65.143759 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
65.07
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
3,821.283118 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
65.106581 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
65.171477 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
65.11
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
3,828.832529 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
54.437075 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
54.569267 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
54.44
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Performance / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 32K · short-answer speed
Median cold document-processing and answer-generation rates; both tasks pooled.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects. Answers capped at 48 tokens. Cold samples are documents, not all scored questions.
8 results
Median per-request generation tok/s from accepted short answers; not sustained decode · — = not available
KV cache · 32K · short-answer speed. Scores are out of 100; No primary score; named metrics retain their own units. Time basis: Not reported for this profile.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,801.125291 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
124.380043 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
118.458616 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
124.38
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,788.498537 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
116.515397 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
110.20702 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
116.52
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,792.908184 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
114.549178 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
107.212053 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
114.55
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,958.567215 tok/sAccepted launches; task-pooled sample population
cold samples
9 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
124.342554 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
116.677555 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
124.34
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
3,588.096134 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
59.811177 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
59.720452 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
59.81
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
3,500.320104 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
59.548424 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
59.633882 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
59.55
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
3,488.279672 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
59.238232 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
59.318621 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
59.24
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
3,502.82737 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
48.075767 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
48.111998 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
48.08
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Performance / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 64K · short-answer speed
Median cold document-processing and answer-generation rates; both tasks pooled.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects. Answers capped at 48 tokens. Cold samples are documents, not all scored questions.
8 results
Median per-request generation tok/s from accepted short answers; not sustained decode · — = not available
KV cache · 64K · short-answer speed. Scores are out of 100; No primary score; named metrics retain their own units. Time basis: Not reported for this profile.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,395.635705 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
116.296009 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
113.554242 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
116.30
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,333.810174 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
103.466784 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
101.027053 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
103.47
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,334.659137 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
103.198518 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
100.683319 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
103.20
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,457.700391 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
115.519306 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
112.387292 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
115.52
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,998.913284 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
49.696626 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
49.695125 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
49.70
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q8Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
65,536 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,929.193346 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
50.989198 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
51.038554 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
50.99
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,925.141119 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
50.609215 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
50.605195 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
50.61
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
2,938.729678 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
34.149471 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
34.243593 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
34.15
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Performance / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · 128K · short-answer speed
Median cold document-processing and answer-generation rates; both tasks pooled.
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects. Answers capped at 48 tokens. Cold samples are documents, not all scored questions.
8 results
Median per-request generation tok/s from accepted short answers; not sustained decode · — = not available
KV cache · 128K · short-answer speed. Scores are out of 100; No primary score; named metrics retain their own units. Time basis: Not reported for this profile.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
1,813.749224 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
112.166097 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
110.057342 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
112.17
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
1,661.617544 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
83.247055 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
81.335467 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
83.25
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
1,666.872276 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
83.746273 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
81.307662 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
83.75
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
1,734.381684 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
92.224273 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
90.17669 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
92.22
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · F16Off · F16Captured: Not recordedCapture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Not recorded
Date precision
No accepted captures for this configuration.
Capture timezone
Not recorded
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
No accepted allocation
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
1,853.885612 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
39.568994 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
39.60688 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
39.57
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · Q4Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
1,852.564122 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
38.688711 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
38.699896 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
38.69
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
Mistral Small 3.2 · K8V4Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Range of accepted question request envelopes; intervals can include downtime.
Capture started
Capture finished
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Up to 48 generated tokens per answer; document primers/cold requests used for cold rate
All measurements and denominators (5)
cold prompt tps median
1,857.849536 tok/sAccepted launches; task-pooled sample population
cold samples
8 samplesAccepted launches; task-pooled sample population
generation samples
200 samplesAccepted launches; task-pooled sample population
generation tps median
21.469661 tok/sAccepted launches; task-pooled sample population
generation tps time weighted
21.543995 tok/sAccepted launches; task-pooled sample population
Saved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
—
21.47
—
—
CompleteSaved accepted results. Within-model questions are paired across cache settings; model tokenizers differ. Context allocations and boot sessions vary, so speed differences are descriptive, not isolated cache effects.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Performance / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · practical context capacity
practical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
practical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K Guard margin at least 1,640 MiB. Allocation and one short-input speed check do not establish full-length speed or accuracy.
8 results
Not measured in this profile · — = not available
KV cache · practical context capacity. Scores are out of 100; No primary score; named metrics retain their own units. Time basis: Not reported for this profile.
No additional benchmark answers; accuracy tested only through 128K
All measurements and denominators (10)
verified context tokens
131,072 tokenspractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
trained context tokens
262,144 tokensGGUF trained-context cap
actual cuda buffers mib
30,503.41 MiBLogged CUDA buffers; excludes baseline desktop usage
baseline max mib
428 MiBUnloaded GPU baseline maximum before this launch
actual headroom mib
1,675.59 MiBPhysical VRAM minus actual logged CUDA buffers and baseline maximum
after load mib
31,438 MiBWhole GPU memory after loading
peak 60s mib
31,486 MiBWhole GPU sampled during the verification window
cold 4k tps
2,785.222928 tok/sOne ungraded 4K request at the larger allocation
speed reference ratio
0.983375 ratio4K cold rate divided by accepted model reference
minimum headroom mib
1,640 MiBRequired startup guard margin
practical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
—
—
—
29.79 GiBLogged CUDA buffers at verified context; excludes desktop baseline
Capacity verifiedpractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
No additional benchmark answers; accuracy tested only through 128K
All measurements and denominators (10)
verified context tokens
208,896 tokenspractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
trained context tokens
262,144 tokensGGUF trained-context cap
actual cuda buffers mib
30,477.68 MiBLogged CUDA buffers; excludes baseline desktop usage
baseline max mib
391 MiBUnloaded GPU baseline maximum before this launch
actual headroom mib
1,738.32 MiBPhysical VRAM minus actual logged CUDA buffers and baseline maximum
after load mib
31,372 MiBWhole GPU memory after loading
peak 60s mib
31,450 MiBWhole GPU sampled during the verification window
cold 4k tps
2,403.222456 tok/sOne ungraded 4K request at the larger allocation
speed reference ratio
0.848503 ratio4K cold rate divided by accepted model reference
minimum headroom mib
1,640 MiBRequired startup guard margin
practical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
—
—
—
29.76 GiBLogged CUDA buffers at verified context; excludes desktop baseline
Capacity verifiedpractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
No additional benchmark answers; accuracy tested only through 128K
All measurements and denominators (10)
verified context tokens
262,144 tokenspractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
trained context tokens
262,144 tokensGGUF trained-context cap
actual cuda buffers mib
28,669.68 MiBLogged CUDA buffers; excludes baseline desktop usage
baseline max mib
428 MiBUnloaded GPU baseline maximum before this launch
actual headroom mib
3,509.32 MiBPhysical VRAM minus actual logged CUDA buffers and baseline maximum
after load mib
29,604 MiBWhole GPU memory after loading
peak 60s mib
29,652 MiBWhole GPU sampled during the verification window
cold 4k tps
2,694.722117 tok/sOne ungraded 4K request at the larger allocation
speed reference ratio
0.951422 ratio4K cold rate divided by accepted model reference
minimum headroom mib
1,640 MiBRequired startup guard margin
practical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
—
—
—
28.00 GiBLogged CUDA buffers at verified context; excludes desktop baseline
Capacity verifiedpractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
No additional benchmark answers; accuracy tested only through 128K
All measurements and denominators (10)
verified context tokens
253,952 tokenspractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
trained context tokens
262,144 tokensGGUF trained-context cap
actual cuda buffers mib
30,429.68 MiBLogged CUDA buffers; excludes baseline desktop usage
baseline max mib
418 MiBUnloaded GPU baseline maximum before this launch
actual headroom mib
1,759.32 MiBPhysical VRAM minus actual logged CUDA buffers and baseline maximum
after load mib
31,383 MiBWhole GPU memory after loading
peak 60s mib
31,431 MiBWhole GPU sampled during the verification window
cold 4k tps
2,473.738001 tok/sOne ungraded 4K request at the larger allocation
speed reference ratio
0.8734 ratio4K cold rate divided by accepted model reference
minimum headroom mib
1,640 MiBRequired startup guard margin
practical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
—
—
—
29.72 GiBLogged CUDA buffers at verified context; excludes desktop baseline
Capacity verifiedpractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
Mistral Small 3.2 · F16Off · F16Campaign captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Mistral Small 3.2 24B Instruct 2506 Q6_K · F16
Test captured
Date precision
Recorded capacity verification window start; not a full-length inference timestamp
Short-input capacity verification window
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
77,824 tokens
K / V cache
F16
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
No additional benchmark answers; accuracy tested only through 128K
All measurements and denominators (10)
verified context tokens
77,824 tokenspractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
trained context tokens
131,072 tokensGGUF trained-context cap
actual cuda buffers mib
30,385.22 MiBLogged CUDA buffers; excludes baseline desktop usage
baseline max mib
428 MiBUnloaded GPU baseline maximum before this launch
actual headroom mib
1,793.78 MiBPhysical VRAM minus actual logged CUDA buffers and baseline maximum
after load mib
31,314 MiBWhole GPU memory after loading
peak 60s mib
31,332 MiBWhole GPU sampled during the verification window
cold 4k tps
4,242.956025 tok/sOne ungraded 4K request at the larger allocation
speed reference ratio
0.993047 ratio4K cold rate divided by accepted model reference
minimum headroom mib
1,640 MiBRequired startup guard margin
practical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
—
—
—
29.67 GiBLogged CUDA buffers at verified context; excludes desktop baseline
Capacity verifiedpractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
Mistral Small 3.2 · Q8Off · Q8Campaign captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8
Test captured
Date precision
Recorded capacity verification window start; not a full-length inference timestamp
Short-input capacity verification window
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q8
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
No additional benchmark answers; accuracy tested only through 128K
All measurements and denominators (10)
verified context tokens
131,072 tokenspractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
trained context tokens
131,072 tokensGGUF trained-context cap
actual cuda buffers mib
29,485.3 MiBLogged CUDA buffers; excludes baseline desktop usage
baseline max mib
428 MiBUnloaded GPU baseline maximum before this launch
actual headroom mib
2,693.7 MiBPhysical VRAM minus actual logged CUDA buffers and baseline maximum
after load mib
30,416 MiBWhole GPU memory after loading
peak 60s mib
30,437 MiBWhole GPU sampled during the verification window
cold 4k tps
4,185.497879 tok/sOne ungraded 4K request at the larger allocation
speed reference ratio
0.979599 ratio4K cold rate divided by accepted model reference
minimum headroom mib
1,640 MiBRequired startup guard margin
practical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
—
—
—
28.79 GiBLogged CUDA buffers at verified context; excludes desktop baseline
Capacity verifiedpractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
Mistral Small 3.2 · Q4Off · Q4Campaign captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4
Test captured
Date precision
Recorded capacity verification window start; not a full-length inference timestamp
Short-input capacity verification window
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
Q4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
No additional benchmark answers; accuracy tested only through 128K
All measurements and denominators (10)
verified context tokens
131,072 tokenspractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
trained context tokens
131,072 tokensGGUF trained-context cap
actual cuda buffers mib
24,365.3 MiBLogged CUDA buffers; excludes baseline desktop usage
baseline max mib
428 MiBUnloaded GPU baseline maximum before this launch
actual headroom mib
7,813.7 MiBPhysical VRAM minus actual logged CUDA buffers and baseline maximum
after load mib
25,296 MiBWhole GPU memory after loading
peak 60s mib
25,314 MiBWhole GPU sampled during the verification window
cold 4k tps
4,180.993923 tok/sOne ungraded 4K request at the larger allocation
speed reference ratio
0.978545 ratio4K cold rate divided by accepted model reference
minimum headroom mib
1,640 MiBRequired startup guard margin
practical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
—
—
—
23.79 GiBLogged CUDA buffers at verified context; excludes desktop baseline
Capacity verifiedpractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
Mistral Small 3.2 · K8V4Off · K8V4Campaign captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4
Test captured
Date precision
Recorded capacity verification window start; not a full-length inference timestamp
Short-input capacity verification window
Capture timezone
UTC
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens
K / V cache
K8V4
MTP
Off
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
No additional benchmark answers; accuracy tested only through 128K
All measurements and denominators (10)
verified context tokens
131,072 tokenspractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
trained context tokens
131,072 tokensGGUF trained-context cap
actual cuda buffers mib
26,925.3 MiBLogged CUDA buffers; excludes baseline desktop usage
baseline max mib
428 MiBUnloaded GPU baseline maximum before this launch
actual headroom mib
5,253.7 MiBPhysical VRAM minus actual logged CUDA buffers and baseline maximum
after load mib
27,856 MiBWhole GPU memory after loading
peak 60s mib
27,877 MiBWhole GPU sampled during the verification window
cold 4k tps
4,183.121768 tok/sOne ungraded 4K request at the larger allocation
speed reference ratio
0.979043 ratio4K cold rate divided by accepted model reference
minimum headroom mib
1,640 MiBRequired startup guard margin
practical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
—
—
—
26.29 GiBLogged CUDA buffers at verified context; excludes desktop baseline
Capacity verifiedpractical max under this setup (display-attached, fresh-reboot baseline); allocation + 4K speed verified; long-context accuracy not tested beyond 128K
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Performance / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · main-cache allocation at 128K
Main KV cache requirements from verified per-token sizes; excludes weights, compute, recurrent state and Qwen draft cache.
The Mistral F16 requirement is unavailable as an accepted 128K run. Requirements are not whole-GPU peaks or a fit guarantee.
8 results
Not measured in this profile · — = not available
KV cache · main-cache allocation at 128K. Scores are out of 100; No primary score; named metrics retain their own units. Time basis: Not reported for this profile.
Model & configuration
Score / 100
Generation tok/s
Not reported for this profile
Memory see basis
Evidence
Qwen3.8-27B · F16Off · F16Captured: Not recordedCapture, runtime & configuration for Qwen3.8-27B UD-Q6_K · F16
Qwen3.8-27B UD-Q6_K · F16
Test captured
Not recorded
Date precision
Derived allocation requirement, not an independent inference capture
Capture timezone
Not recorded
Application / version
llama-server b11580
Inference engine / build
llama.cpp b11580 (4413a1924)
OS / environment
Windows; RTX 5090 32 GB, display-attached as reported by Casey; Intel i9-13900K; 128 GB DDR5; benchmark-era Windows build not recorded
GPU driver version
617.14
CUDA runtime version
Windows CUDA 13.4 build
Thinking
Off
Sampling
Greedy; temperature 0
Output cap
48 tokens
Flash attention
Enabled
GPU offload
Full for accepted runs
Batch / microbatch
2048 / 512
Parallel requests
1
Automatic fit / context shifting
Disabled
Context capacity
131,072 tokens (requirement calculation)
K / V cache
F16
MTP
On, maximum 3 draft tokens; draft KV F16
Prompt reuse
One document prefix reused across questions; Qwen primed before questions
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
68 results
Not measured in this profile · — = not available
KV cache · paired accuracy and answer churn. Scores are out of 100; No primary score; named metrics retain their own units. Time basis: Not reported for this profile.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q8 · vs-f16-4k-assetOff · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16-4k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-f16-4k-assetOff · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16-4k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-f16-4k-assetOff · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16-4k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q8 · vs-f16-4k-configOff · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16-4k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-f16-4k-configOff · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16-4k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-f16-4k-configOff · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16-4k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q8 · vs-f16-16k-assetOff · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16-16k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-f16-16k-assetOff · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16-16k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-f16-16k-assetOff · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16-16k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q8 · vs-f16-16k-configOff · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16-16k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-f16-16k-configOff · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16-16k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-f16-16k-configOff · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16-16k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q8 · vs-f16-32k-assetOff · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16-32k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-f16-32k-assetOff · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16-32k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-f16-32k-assetOff · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16-32k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q8 · vs-f16-32k-configOff · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16-32k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-f16-32k-configOff · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16-32k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-f16-32k-configOff · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16-32k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q8 · vs-f16-64k-assetOff · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16-64k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-f16-64k-assetOff · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16-64k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-f16-64k-assetOff · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16-64k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q8 · vs-f16-64k-configOff · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16-64k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-f16-64k-configOff · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16-64k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-f16-64k-configOff · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16-64k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q8 · vs-f16-128k-assetOff · Q8Captured: Not recordedCapture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16-128k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
UnavailableUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-f16-128k-assetOff · Q4Captured: Not recordedCapture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16-128k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
UnavailableUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-f16-128k-assetOff · K8V4Captured: Not recordedCapture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16-128k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
UnavailableUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-q8-128k-assetOff · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-q8-128k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-q8-128k-assetOff · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-q8-128k-asset
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q8 · vs-f16-128k-configOff · Q8Captured: Not recordedCapture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16-128k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
UnavailableUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-f16-128k-configOff · Q4Captured: Not recordedCapture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16-128k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
UnavailableUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-f16-128k-configOff · K8V4Captured: Not recordedCapture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16-128k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
UnavailableUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · Q4 · vs-q8-128k-configOff · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-q8-128k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
Mistral Small 3.2 · K8V4 · vs-q8-128k-configOff · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-q8-128k-config
Unavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
—
—
—
—
Paired comparisonUnavailable Mistral F16/128K comparisons remain null. Paired question identity is within model. The 62-test family contains F16 comparisons and Q8 comparisons at 128K; no equivalence claim.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
KV cache / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · pooled differences versus F16
Question-weighted paired correctness differences with 95% whole-document cluster-bootstrap intervals.
10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
6 results
Not measured in this profile · — = not available
KV cache · pooled differences versus F16. Scores are out of 100; No primary score; named metrics retain their own units. Time basis: Not reported for this profile.
10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
—
—
—
—
Pooled comparison10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
—
—
—
—
Pooled comparison10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
—
—
—
—
Pooled comparison10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
Mistral Small 3.2 · Q8 · vs-f16Off · Q8Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q8 · vs-f16
10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
—
—
—
—
Pooled comparison10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
Mistral Small 3.2 · Q4 · vs-f16Off · Q4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · Q4 · vs-f16
10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
—
—
—
—
Pooled comparison10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
Mistral Small 3.2 · K8V4 · vs-f16Off · K8V4Captured: Capture, runtime & configuration for Mistral Small 3.2 24B Instruct 2506 Q6_K · K8V4 · vs-f16
10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
—
—
—
—
Pooled comparison10,000 resamples. Intervals include zero; this is no detectable difference, not equivalence or a universal effect bound. Long-context task cells have four documents.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Performance / Reviewed: 5 Nov 2026; capture dates are per result
KV cache · Qwen MTP on/off
Qwen F16 paired short-answer MTP comparison: 200 questions, no parsed or byte differences.
MTP on allocated 65,536 tokens; off allocated 36,864, and boots differ. Observed 2.24× ratio is not an isolated causal speedup or universal answer-preservation guarantee.
2 results
Median per-request generation tok/s over 200 short answers · — = not available
KV cache · Qwen MTP on/off. Scores are out of 100; No primary score; named metrics retain their own units. Time basis: Not reported for this profile.
2.237247 ratioDifferent boot and allocation; descriptive comparison
allocated context tokens
65,536 tokensActual allocated capacity
generation tps time weighted
118.458616 tok/sTotal generated tokens divided by total generation time
MTP on allocated 65,536 tokens; off allocated 36,864, and boots differ. Observed 2.24× ratio is not an isolated causal speedup or universal answer-preservation guarantee.
—
124.38
—
—
Ablation comparisonMTP on allocated 65,536 tokens; off allocated 36,864, and boots differ. Observed 2.24× ratio is not an isolated causal speedup or universal answer-preservation guarantee.
2.237247 ratioDifferent boot and allocation; descriptive comparison
allocated context tokens
36,864 tokensActual allocated capacity
generation tps time weighted
55.473492 tok/sTotal generated tokens divided by total generation time
MTP on allocated 65,536 tokens; off allocated 36,864, and boots differ. Observed 2.24× ratio is not an isolated causal speedup or universal answer-preservation guarantee.
—
55.60
—
—
Ablation comparisonMTP on allocated 65,536 tokens; off allocated 36,864, and boots differ. Observed 2.24× ratio is not an isolated causal speedup or universal answer-preservation guarantee.
No matching results in this campaign.
Try a shorter model name or choose another campaign.
Records include repeated configurations, evaluator revisions, and incomplete attempts. They are not a count of unique models or independent reruns.
SMALLER GPU, CLEARER EXPECTATIONS
12 / 16 GB fit screening.
Projected · RTX 5090 measured
11 files from the reconciled comparison. Values are the highest available dedicated-process samples at a 64K allocation with Q4 K/V across QuantBench and context probes. They do not establish quality rankings or physical target-card fit.
Candidate capture dates & runtimes: Dates, versions, and settings for all 11 candidates appear below. The 11 process-memory values were measured on RTX 5090; smaller-card fit is projected.
Workload estimate + 2 GiB reserveScreen against 12 GiB and 16 GiB budgets. Physical smaller-card testing remains unverified.
Projected memory screening with an assumed 2 GiB desktop and application reserve
Model / weight quant
Measured process peak
With reserve
12 GiB budget
16 GiB budget
Qwen3.8-27B Heretic GSQ-RCOIQ3_XXS
11.34 GiB
13.34 GiB
Above estimate
Within estimate
Ternary Bonsai 2 27BPTQ1_0
7.48 GiB
9.48 GiB
Within estimate
Within estimate
Ternary Bonsai 2 27BPQ2_0
8.59 GiB
10.59 GiB
Within estimate
Within estimate
Gemma 4 12B IT QATQ4_0
7.91 GiB
9.91 GiB
Within estimate
Within estimate
Qwen3.5 9BQ5_K_M
8.84 GiB
10.84 GiB
Within estimate
Within estimate
Qwen3.5 4BQ5_K_M
4.35 GiB
6.35 GiB
Within estimate
Within estimate
Ornith 1.5 9BQ5_K_M
7.90 GiB
9.90 GiB
Within estimate
Within estimate
Qwen3 4B Instruct 2507Q5_K_M
6.01 GiB
8.01 GiB
Within estimate
Within estimate
Ministral 3 14B Instruct 2512Q4_K_M
10.93 GiB
12.93 GiB
Above estimate
Within estimate
Ministral 3 8B Instruct 2512Q4_K_M
7.75 GiB
9.75 GiB
Within estimate
Within estimate
Ministral 3 3B Instruct 2512Q5_K_M
4.92 GiB
6.92 GiB
Within estimate
Within estimate
Approximately two-second Windows DedicatedUsage samples exclude other processes and shared memory and can miss short transients. The reserve is a planning choice. Configured full offload does not prove actual layer residency.
Capture dates, runtimes & settings for all 11 candidates
These dates describe the selected process-memory samples. Quality profiles carry their own capture dates and configurations. Missing settings stay unavailable.
Exact recorded GGUF content; completion status is unchanged
Other context sizes and quality
The memory profile includes measured 32K/64K probes and separately labeled 128K extrapolations where available. The article also separates Bonsai Q4 and Q8 memory at a 262,144-token allocation. Successful allocation does not establish long-context answer quality.
READ THE CONDITIONS FIRST
A score needs its settings.
01
Compare within a campaign.
QuantBench versions use different tasks and scoring rules. HomHaystack versions use different lengths and weights. A 94 in one profile does not outrank an 83 in another.
02
Capacity is not input length.
A loaded 64K window does not prove a 64K prompt was tested. Weight quant, K/V precision, thinking, and MTP all affect the result and memory use.
03
Speed is only part of the wait.
Generation rates exclude prompt processing. Mean and median rates stay labeled. Suite duration is separate from time to first token and time to a usable answer.
04
Fit estimates have limits.
Whole-board peak, peak above baseline, and projected use are different measures. GB and GiB remain distinct. RTX 5090 speed is not a prediction for a smaller GPU.
Blender structural /8 checks and subjective editorial /10 ratings have their own scales. Time may mean a suite, an individual attempt or a mean attempt; retain its label. Unavailable results stay unavailable. Partial, timed-out, unsupported, and regraded entries retain their status. This explorer does not calculate a cross-campaign or combined quality score.
How capture dates and runtime versions are recorded
A campaign date applies to its grouped tests and attempts; it does not imply an exact timestamp or successful completion for every row. Exact task timestamps include the original UTC offset when available. Regraded results retain the original capture metadata because rescoring saved answers is not new inference.
Application versions and inference-engine builds are separate: an LM Studio version does not establish its llama.cpp build. OS, driver, CUDA, and other runtime versions are shown only when tied to the capture. “Not recorded” means the available source does not establish that field; current machine settings and requested retest settings are not substituted.
Captured September 25 CDT: 50/50 QuantBench requests and 9/9 Legacy HomHaystack requests. LM Studio 0.4.25 Build 1 with CUDA12 llama.cpp extension 2.43.0 selected. The workbook includes task start/finish, verified settings, task-scoped memory, and median generation rates. The upstream engine commit is not recorded.
NVFP4 vs GGUF: current results and repair
Nine configurations, 18 current result records: 448 completed QuantBench questions and 80 completed reduced HomHaystack requests. ThinkingCap GGUF has no QuantBench overall at 48/50; OrcaRouter GGUF has no HomHaystack overall at 8/9.
Repaired vLLM 0.29.0 results are separate from the original vLLM 0.27.1 speed baseline. ThinkingCap NVFP4 completed both suites after repair. All generation rates exclude prefill; effective request rates remain separate.
Eleven measured process-memory peaks and projected 12/16 GB screenings are listed above. The new article links 71 records across 11 profiles, including both thinking modes, original/adjusted Instruct cohorts, separate Q4/Q8 Haystack components and memory.
Benchmark editions and comparison coverage
Benchmark editions use different tasks and scoring rules. Original video-era scores and later saved-answer regrades have distinct profiles. The original Q2–Q5 Haystack records include all 12 final Reasoning-v2 start times; CDT is assumed for timestamps recorded without a UTC offset.
Some ISTA results also appear in the uncensored campaign tables, with the ISTA article linked for additional analysis. Publisher benchmarks and the qualitative Opus/Sol website demonstration are covered in the frontier-model article.
I want to know whether a model can do useful work on hardware I can actually run.
QuantBench helps me see which tasks it handles and where it falls apart. HomHaystack checks whether it can find and use details in a long context. I look at both, along with memory use and how long I’m waiting for an answer.
The setup matters. I keep the quant, thinking mode, context, cache, and benchmark edition with the result. An unfinished run stays unfinished. A projected fit on a smaller GPU stays a projection.
The scores give me a shortlist. Then I try the models on the work I actually need done.
The recommendations weigh QuantBench, HomHaystack, and the recorded memory use. Most quality testing was on my RTX 5090. Guidance for 12, 16, and 24 GB cards is a starting point to verify on your own hardware.
Practical evaluations for people working with local AI, hardware, and cybersecurity.
I’m interested in product evaluations, sponsored content, and technical collaborations that are relevant to the work I cover here.
Email me or contact me on X with a brief overview of your product, the proposed scope, and timing. Sponsorships and loaned hardware will be disclosed, and my conclusions will remain independent.
