A tuning method cuts Gemma's thinking tokens by a third at the same accuracy

BottleCap AI published results for ThinkingCap, a method that shortens a reasoning model's thinking phase, applied here to Google's gemma-4-26B-A4B, a sparse mixture-of-experts model with about 4B active parameters per token. The goal is narrow: reach the same answer with fewer thinking tokens and change nothing else, not knowledge, answer quality, instruction following, or safety.

On BottleCap's own evaluation across twelve benchmarks, the tuned model used 38.9% fewer thinking tokens on average, and up to 69.1% fewer on LongBench v2. Accuracy barely moved: 75.03% versus 75.18% for the original, a difference of 0.15 points, with five benchmarks actually scoring higher and the worst case down 4 points on AA-LCR. Two reliability problems also dropped sharply. Responses flagged as looping fell from 3.64% to 0.76%, and responses whose thinking never closed fell from 4.35% to 0.72%. BottleCap ran each benchmark with multiple seeds on the same harness and notes that most of the accuracy differences are within run-to-run noise.

These are vendor-run numbers, so treat them as a claim to verify rather than a settled result. The company also mentions a PRO version with larger reductions but gives no evaluation data for it.

Why it matters

If you serve a reasoning model in production, thinking tokens are a direct cost and latency line. A drop-in model that thinks about 39% less for the same answers would cut both, so this is worth benchmarking on your own traffic once weights are available, keeping in mind the numbers are self-reported.