BottleCap Slashes Reasoning Tokens by 37%

🔥 What happened BottleCap AI released ThinkingCap-Qwen3.8-27B, a fine-tune of Qwen's 27B model that aggressively cuts reasoning tokens. Across 12 benchmarks, it uses 37.2% fewer thinking tokens on average, while accuracy drops just 0.86 percentage points. 💡 Why it matters On MMMLU, thinking tokens shrink 65.5% – from 1,656 to 571 – with virtually no accuracy loss. On long-context tasks (AA-LCR), accuracy actually rises 2.25 points. It's a drop-in replacement for Qwen3.8-27B on vLLM or SGLang. ⚡ Our take If you deploy reasoning models, you pay per token. 37% less thinking at near-identical performance isn't a benchmark trick – it's a direct cost lever. The license is restrictive, though: commercial use requires a deal with BottleCap.