BottlecapAI cuts Gemma-4-26B-A4B's thinking by up to 69% with almost no accuracy loss
Gemma-4-26B-A4B with 39-69% less thinking at same accuracy
BottlecapAI's fifth ThinkingCap model trims how long Google's gemma-4-26B-A4B-it reasons before answering: thinking tokens drop 38.9% on average and up to 69.1% across twelve benchmarks, at a cost of just 0.15pp of accuracy. Looping falls from 3.64% of responses to 0.76%, and it's a drop-in replacement on the same vLLM stack with unchanged sampling settings.
The goal is to spend fewer tokens getting to the answer, and change nothing else: preserve the knowledge, reasoning ability, answer quality, instruction following and safety behaviour.