ThinkingCap cuts Gemma 4 reasoning tokens with minimal accuracy loss
Bottlecap AI released ThinkingCap-gemma-4-26B-A4B-it, a drop-in replacement designed to reduce the amount of thinking performed by Google's Gemma 4 26B-A4B-it. In its own evaluation across 12 benchmarks, the model used 38.9% fewer thinking tokens on average while average accuracy fell by 0.15 percentage points, although AA-LCR accuracy dropped by 4 points. Bottlecap also reports lower looping and published the model with GGUF builds.
The release offers a documented approach to reducing inference-time reasoning tokens while largely preserving Gemma 4 accuracy, though the benchmark results are vendor-reported and include a notable AA-LCR regression.
TL;DR
This is the fifth model in the ThinkingCap series. We reduced how much Google's gemma-4-26B-A4B-it thinks, without compromising the quality of its answers.
- Up to 69.1% reduction in thinking tokens, on average 38.9% measured across twelve benchmarks, for 0.15pp of accuracy
- The first ThinkingCap on a Google model, and the smallest accuracy cost of any ThinkingCap so far
- Looping drops from 3.6% of the original model's responses to 0.8% of ours
- Same sampling settings (thinking on, temperature 1.0, top_p 0.95), same serving stack (vLLM), drop-in replacement
ThinkingCap PRO: want even better efficiency? The PRO version of this model uses up to 87.6% fewer thinking tokens, 62% on average. Faster answers and lower inference costs at scale. Contact us at [email protected].
Our fifth ThinkingCap
We have published four token-efficient models so far: ThinkingCap-Qwen3.6-27B, ThinkingCap-Qwen3.8-27B, ThinkingCap-NVIDIA-Nemotron-3.5-Lightning-30B-A3B and ThinkingCap: Qwen3.8-Flash-Next.
#Bottlecap AI#Google#ThinkingCap#Gemma 4 26B-A4B-it#vLLM#GGUF