Gemma-4-26B-A4B with 39-69% less thinking at same accuracy

BottleCap AI reports a new ThinkingCap variant applied to Google's sparse Mixture-of-Experts model gemma-4-26B-A4B-it that sharply reduces "thinking" tokens while keeping accuracy essentially unchanged. The tuned model cuts mean reasoning-token usage by 38.9% across twelve benchmarks (8773→5398 mean tokens) and achieves up to 69.1% reduction on LongBench v2, with an average accuracy change of −0.15 percentage points. Some task families even improve (instruction following +0.65pp, maths +0.39pp), while the largest accuracy drop is −4.00pp on AA-LCR; code performance is effectively flat. Looping behavior falls from 3.64% of responses to 0.76%, and truncation rates decline in many cases. A PRO variant promises up to 87.6% thinking-token reduction (62% on average) for greater speed and cost savings.

Evaluation used identical serving and sampling settings (vLLM, thinking on, temperature 1.0, top_p 0.95, up to 65,536 token context) with multiple seeds per benchmark and standard scoring (exact match, test-suite runs for code, human-judged tasks where applicable). Results are reported as means with 95% confidence intervals across seeds; detailed per-benchmark numbers show consistent token reductions across long-context, knowledge, reasoning, maths, code, and instruction families. The model is published as bottlecapai/ThinkingCap-gemma-4-26B-A4B-it with GGUF builds and is presented as a drop-in replacement.