Sakhanda Wire
NVDA $154.00 -1.37% MSFT $506.28 -2.26% GOOGL $280.87 -1.00% META $620.34 -0.97% AMZN $256.88 -0.34%
← Back to the news

ThinkingCap-Qwen3.8-27B Cuts Reasoning Tokens by 37.2%

BottleCap AI has released ThinkingCap-Qwen3.8-27B, the second model in its ThinkingCap series. The fine-tune of Qwen3.8-27B was designed to produce shorter reasoning traces without intentionally changing the model’s knowledge, instruction following or safety behaviour. Across 12 benchmarks, thinking-token use fell by an average of 37.2%, while macro accuracy declined from 86.65% to 85.79%, a reduction of 0.86 percentage points.

Results varied by task. AA-LCR long-context accuracy increased by 2.25 percentage points, reaching 84.00% with 38.6% fewer thinking tokens. LiveCodeBench v6 rose by 0.07 points, while AIME 2026 fell by 3.85 points to 94.27%. Under a 16K-token response limit, the new model scored higher than the base model, and the rate of truncated traces decreased from 0.51% to 0.34%.

The model is intended as a drop-in replacement on vLLM and SGLang and is available in FP8, NVFP4, GGUF and MLX builds. Evaluation used one NVIDIA H200 with identical sampling settings. The gated weights use the PolyForm Small Business 1.0.0 license alongside a BottleCap personal-use grant; commercial use beyond the small-business terms requires an agreement with BottleCap.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news