ThinkingCap Qwen3.8-27B

Hard science questions. This is a science reasoning test, not an overall intelligence score.

Benchmark efficiency: 4.994 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.

Test settings: Full 198-question GPQA Diamond, four seeds, selected Q4_K_M on llama.cpp bfd73a8 and RTX PRO 6000 Max-Q with eight parallel slots; --jinja, xhigh reasoning, temperature 1.0, top-p 0.95, top-k 20, min-p 0 and a 65,536-token generation cap. Displayed accuracy is the publisher's rounded BF16 reference, 88.0%, plus its reported Q4 delta of -0.9 percentage points.

Publisher-run artifact evaluation, not independent validation; the Q4 and BF16 engines and GPUs differ. The 17,442,400,352-byte language file excludes runtime, KV cache and context memory, and multimodal use adds a separate 931,146,272-byte F16 projector. File size alone does not prove device speed or compatibility. The embedded MTP head needs the publisher's documented current llama.cpp architecture support; older builds can fail to load. PolyForm Small Business 1.0.0 plus BottleCap's personal-use grant applies; organizations outside the small-business terms need a separate commercial license. Sampled decoding is recommended because the publisher warns greedy decoding can loop.

Publisher Q4_K_M evaluation on a fixed 1,500-question MMLU-Pro subset. Kept separate from full-test MMLU-Pro results.

Benchmark efficiency: 4.867 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.

Test settings: Fixed 1,500-question MMLU-Pro subset, one seed, selected Q4_K_M on llama.cpp bfd73a8 and RTX PRO 6000 Max-Q with eight parallel slots; the same sampled xhigh settings and 65,536-token cap as the card's quantization study. Displayed accuracy is the 84.1% rounded BF16 reference plus the reported +0.8-point Q4 delta.

Publisher-run subset result, not the full MMLU-Pro test or independent validation. The card describes the Q4 result as evaluated rather than lossless; engine, GPU, sampling and rounding can affect the comparison.

Publisher Q4_K_M evaluation on 300 IFBench prompts over two seeds. The card does not identify this as IFEval prompt-strict, so it remains separate.

Benchmark efficiency: 4.609 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.

Test settings: IFBench on 300 prompts over two seeds, selected Q4_K_M on llama.cpp bfd73a8 and RTX PRO 6000 Max-Q with eight parallel slots; --jinja, xhigh reasoning and the card's sampled decoding settings. Displayed accuracy is the 79.7% rounded BF16 reference plus the reported +0.7-point Q4 delta.

Publisher-run limited sample, not independent validation and not substituted for the catalog's IFEval prompt-strict metric. Engine, GPU, sampling and rounded-reference arithmetic can affect the result.

Publisher Q4_K_M evaluation on 100 AA-LCR items, graded by Gemma-4-26B-A4B-it with thinking disabled. Specialist long-context result.

Benchmark efficiency: 4.472 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.

Test settings: AA-LCR on 100 items, one seed, selected Q4_K_M on llama.cpp bfd73a8 and RTX PRO 6000 Max-Q with two parallel slots; graded by Gemma-4-26B-A4B-it with thinking disabled. Displayed accuracy is the 81.0% rounded BF16 reference plus the reported -3.0-point Q4 delta.

Small, publisher-run long-context result with a model judge, not independent validation. Judge choice, parallelism, context and rounded-reference arithmetic materially limit comparison.

Publisher multimodal Q4_K_M evaluation on 765 RealWorldQA items over two seeds, with the separate F16 vision projector loaded.

Benchmark efficiency: 4.753 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.

Test settings: RealWorldQA on 765 items over two seeds, selected Q4_K_M plus the F16 vision projector on llama.cpp bfd73a8 and RTX PRO 5000 with four parallel slots; --jinja, xhigh reasoning and the card's sampled decoding settings. Displayed accuracy is the 83.1% rounded BF16 reference plus the reported -0.2-point Q4 delta.

Publisher-run multimodal result, not independent validation. The separate 931,146,272-byte projector is required for image input and is intentionally excluded from the ranked language-weight bytes; runtime, context and KV cache require additional memory.