ThinkingCap: Gemma-4-26B-A4B — up to 69% and on average 39% less thinking, at unchanged accuracy
Summary
BottleCap AI announces ThinkingCap Gemma-4-26B-A4B, a token-efficient variant of Google's gemma-4-26B-A4B-it that achieves up to 69.1% fewer thinking tokens with an average of 38.9% reduction while maintaining 0.15 percentage points of accuracy. The post details the training objective, evaluation across twelve benchmarks, and comparisons to the base model, including a call to action for ThinkingCap PRO.