BottleCap AI has released ThinkingCap-Qwen3.8-27B, the second model in its ThinkingCap series. It is a fine-tune of the Qwen team’s Qwen3.8-27B with one narrow goal: shorter reasoning traces. Across 12 benchmarks, it spends 37.2% fewer thinking tokens on average. Macro-average accuracy moves from 86.65% to 85.79%, a 0.86pp drop.
Deployable? Yes. It drops in for Qwen3.8-27B on vLLM or SGLang, with FP8, NVFP4, GGUF and MLX builds. The repo is gated, and commercial use beyond the small-business license needs a BottleCap agreement.
What Problem Does ThinkingCap Target?
Reasoning models often spend more thinking tokens than a question needs. BottleCap’s position is that many of those extra tokens do not change the final answer. The first release in the series applied this idea to Qwen3.6-27B.
The objective this time was deliberately conservative. BottleCap did not try to add knowledge or change answer style. Reasoning ability, instruction following and safety behaviour were meant to pass through untouched. The research team also focused harder on math, reasoning, long-context and agentic benchmarks.
Benchmark Results at xhigh Effort
All main numbers use reasoning_effort=xhigh, the chat template default. Every benchmark gets shorter, with cuts ranging from 10.7% to 65.5%.
Knowledge and multilingual tasks shrink the most. MMMLU drops 65.5% (1,656 to 571 tokens) and MMLU-Pro drops 57.3%. GPQA-Diamond falls from 12,772 to 7,267 tokens, a 43.1% cut. IFBench thinks 46.4% less with accuracy nearly flat (79.75% to 79.71%).
Long-context retrieval improves. AA-LCR accuracy rises 2.25pp, from 81.75% to 84.00%, with 38.6% fewer thinking tokens. LiveCodeBench v6 edges up 0.07pp while thinking 20.3% less.
Agentic results hold close to the base. τ²-bench gives up 1.01pp for a 30.9% cut. Terminal-Bench 2.1 loses 0.56pp, well inside its ±4.26 interval, for a 10.7% cut.
The most expensive trade is AIME 2026. Accuracy falls 3.85pp, from 98.13% to 94.27%, for 30.2% less thinking.
Please note that the 37.2% figure is the mean of the 12 per-benchmark reductions. Pooled mean thinking tokens fall from 15,735 to 12,144.
BottleCap also reports a budget curve. Under a 16K-token cap per response, ThinkingCap scores higher than the base model. Truncated traces fall from 0.51% to 0.34%, and looping from 0.06% to 0.05%.
Credit: Source link
