BottleCap launches ThinkingCap-Qwen3.8-27B, a fine-tune that cuts reasoning tokens 37.2% on average

BottleCap has released ThinkingCap-Qwen3.8-27B, a fine-tune of Qwen3.8-27B whose sole purpose is to shorten the chain of thought. Across 12 benchmarks the model reduces reasoning tokens by 37.2% on average, at the cost of a 0.86 percentage-point drop in macro accuracy, from 86.65% to 85.79%. The model drops straight into vLLM or SGLang with FP8, NVFP4, GGUF and MLX builds, though the repository is gated and commercial use beyond the small-business license requires a separate agreement with BottleCap.

Fewer tokens, same result

The most aggressive cuts appear on knowledge and multilingual tasks: MMMLU falls 65.5% (from 1,656 tokens to 571) and MMLU-Pro by 57.3%. On GPQA-Diamond reasoning drops from 12,772 to 7,267 tokens, a 43.1% reduction. IFBench accuracy barely moves (79.75% to 79.71%) while reasoning is cut 46.4%. Long-context retrieval actually improves: AA-LCR gains 2.25 percentage points to 84% with 38.6% fewer tokens, and LiveCodeBench v6 adds 0.07 points with 20.3% less reasoning. Agentic tasks remain stable: τ²-bench loses 1.01 points for 30.9% savings, and Terminal-Bench 2.1 falls 0.56 points — within the ±4.26 margin of error — with only a 10.7% cut. The steepest price is on AIME 2026: a 3.85-point drop (from 98.13% to 94.27%) for 30.2% fewer tokens. The weighted average falls from 15,735 tokens to 12,144.

Interaction with the reasoning-effort dial

Qwen3.8-27B exposes a reasoning-effort setting, and BottleCap's compression stacks on top of it. At medium the base model cuts 52.1% of tokens with a 9.16-point loss, while ThinkingCap cuts 60.2% with a 9.90-point loss. At low the base cuts 55.4% with a 9.71-point loss; ThinkingCap cuts 62.3% with a 10.79-point loss. With reasoning disabled entirely, ThinkingCap trails the base by 5.7 points. The team recommends xhigh for the best accuracy-token balance and notes that individual reasoning modes will be addressed in a future release.

Testing methodology

Both models ran through the same harness on a single H200 with vLLM 0.29.0. Sampling parameters were identical: temperature 1.0, top_p 0.95, top_k 20, min_p 0.0. Accuracy is reported as a multi-seed mean with a 95% confidence interval, ranging from 32 seeds on AIME 2026 down to a single seed on MMLU-Pro and MMMLU. The latter uses a fixed 10,000-question sample; the other 11 benchmarks run the full set. Speculative decoding with MTP (3 draft tokens) proved neutral for accuracy on AIME 2026, with a 53% acceptance rate yielding roughly 2.6 tokens per step, identical to the base model.

Deployment and licensing

The bf16 checkpoint weighs 28 billion parameters and accepts both image and text input. BottleCap publishes five quantized builds: FP8 at 31 GB for vLLM on Hopper and Blackwell; NVFP4 weights-only at 21 GB for vLLM on Hopper (Marlin kernel) and Blackwell; NVFP4 W4A4 AWQ at 23 GB for Blackwell only; GGUF in the 16–55 GB range for llama.cpp, LM Studio and Ollama; and MLX 4-bit DWQ at 21 GB for Apple silicon.