AI BriefingBottleCap releases ThinkingCap for Gemma-4-26B-A4B, cutting thinking tokens by 39% on average
·2026.10.23 23:43
Key point
The new model reduces thinking tokens by up to 69.1% with only a 0.15pp average accuracy drop across twelve benchmarks.
Details
- BottleCap AI released ThinkingCap-gemma-4-26B-A4B-it, a fine-tuned version of Google's Gemma-4-26B-A4B that reduces thinking token usage by 38.9% on average without significant accuracy loss.
- Across twelve benchmarks, the model achieves an average accuracy change of -0.15pp, with five benchmarks showing improved performance and the largest drop of 4.00pp on AA-LCR.
- Looping errors decrease from 3.64% in the base model to 0.76% in the ThinkingCap version, with the most significant reductions on math benchmarks like AIME 2026 and HMMT.
- The model is available on Hugging Face with GGUF builds and serves as a drop-in replacement using the same vLLM serving stack and sampling settings as the original.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
MiniMax M3.1 Flash Launches with 70% Cost Reduction and 3x Faster Inference