BottleCap AI Releases ThinkingCap-Qwen3.8-27B: 37.2% Fewer Thinking Tokens at a 0.86pp Accuracy Cost
BottleCap AI released ThinkingCap-Qwen3.8-27B, cutting thinking tokens 37.2% with a 0.86-point accuracy drop.
BottleCap AI released ThinkingCap-Qwen3.8-27B, a fine-tune of Qwen3.8-27B meant to shorten reasoning traces without adding knowledge. Across 12 benchmarks at reasoning effort xhigh, average thinking tokens fell 37.2% while macro accuracy moved from 86.65% to 85.79%. AIME 2026 fell 3.85 points to 94.27%, while AA-LCR rose 2.25 points to 84.00%. The 28B multimodal checkpoint is gated under PolyForm Small Business 1.0.0 and serves as a drop-in on vLLM or SGLang.
- Fine-tune of Qwen3.8-27B aimed only at shorter reasoning traces.
- Macro accuracy fell from 86.65% to 85.79% across 12 benchmarks.
- AIME 2026 dropped 3.85 points; AA-LCR rose 2.25 points.
- Gated under PolyForm Small Business; broader commercial use needs a license.
- Drop-in for Qwen3.8-27B on vLLM or SGLang, with FP8, NVFP4, GGUF, and MLX builds.
- Organizations
- BottleCap AI
From the original article
4 min read at marktechpost.com
BottleCap AI has released ThinkingCap-Qwen3. 8-27B, a fine-tune of Qwen3. 8-27B that spends 37.
By Michal Sutter · marktechpost.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.