ThinkingCap cuts Gemma 4 reasoning tokens by 38.9%

BottleCap AI released ThinkingCap, a retrained Gemma 4 26B-A4B that uses on average 38.9% fewer reasoning tokens while scoring 75.0 versus 75.2 for the base model across 12 benchmarks. Weights and GGUF versions were published on October 23, 2026 under Apache 2.0.

• Average accuracy 75.0 vs 75.2 for Gemma 4 with 38.9% fewer tokens
• Looping responses fell from 3.64% to 0.76%
• AA-LCR accuracy dropped 4.00 points, the main trade-off
• Token savings only partly hold with GGUF quantization