ThinkingCap has released its fifth token-efficient model, built on Google's gemma-4-26B-A4B-it. The goal was to reduce how much the model thinks before answering while changing nothing else, including its knowledge, reasoning, answer quality, instruction following and safety behaviour. Across twelve benchmarks, thinking tokens fell by 38.9% on average, with reductions of up to 69.1% on individual tests.
The accuracy cost was minimal. Averaged across all twelve benchmarks, accuracy changed by just -0.15pp, which the team notes is smaller than the run-to-run spread on most of the suite. Instruction following and agentic tasks gained 0.65pp and maths gained 0.39pp, while code was essentially flat. Knowledge and reasoning lost 0.51pp and long-context and multimodal tasks lost 0.92pp. Looping, where the model repeats itself, dropped from 3.6% of the original model's responses to 0.8%.
The reductions were largest on long-context and multimodal tasks, at 54.4%, followed by knowledge and reasoning at 44.4%, maths at 36.1% and code at 27.6%. Instruction following and agentic tasks saw the smallest change, at 17.2%.
The model is a drop-in replacement and uses the same sampling settings and serving stack. A PRO version reportedly cuts thinking tokens by up to 87.6%, or 62% on average, for enterprises seeking lower inference costs at scale.
Source: Hacker News · Summarized by HeadlinesBriefing