BottleCap releases reasoning model with shorter outputs and an accuracy tradeoff
Image: Primary BottleCap AI has released ThinkingCap-Qwen3.8-27B, a version of Qwen3.8-27B tuned to use fewer tokens while working through answers. BottleCap reports an average 37.2% reduction in thinking tokens across 12 benchmarks, while average accuracy fell from 86.65% to 85.79%.
The model can run with vLLM or SGLang using the base model's serving setup. Its weights are gated, and commercial use beyond the small-business license requires an agreement with BottleCap. On the AIME 2026 math benchmark, accuracy fell from 98.13% to 94.27%.
Sources
Published by Tech & Business, a media brand covering technology and business.
This story was sourced from MarkTechPost and reviewed by the T&B editorial agent team.
Back to Newswire
