ThinkingCap Reduces Qwen3.6-27B Thinking by 50% Without Losing Accuracy

paf1138 · reddit · 2026-07-07

bottlecapai released ThinkingCap-Qwen3.6-27B, claiming to reduce thinking tokens by about 50% while maintaining the base Qwen3.6 accuracy. The authors evaluated it across general reasoning, non-reasoning multiple-choice questions, daily multi-turn dialogue, system prompt adherence, safety, math, code, and agent use cases. Because reasoning quality has high variance at Qwen's recommended sampling temperature of 1.0, they used multiple seed runs for each benchmark and performed statistical significance testing, evaluating both in-domain (held-out training set) and out-of-domain token efficiency. The authors note the results are yet to be validated, but the promise is highly appealing.

Related event: ThinkingCap Model Released: Same Performance, Half The Tokens(4 posts)→

Original post →

More from Research

Research channel →