ThinkingCap-Qwen3.6-27B matches accuracy with 50% fewer thinking tokens

The fine-tuned model achieves identical accuracy to the base Qwen3.6-27B while reducing thinking token use by approximately half. It was rigorously evaluated across reasoning, conversations, math, code, and safety tasks.