Cut Qwen3.8-27B Reasoning Tokens by 40% -- 3.8 'ThinkingCap' benchmarked!

UkisAI's Swift-Qwen3.8-27B model has been benchmarked, showing significant efficiency improvements over the base Qwen3.8-27B. It achieves a comparable success rate (Pass1 30.8% vs 27.1%, Pass2 75.7% vs 77.6%) while using 63% of the tokens (12.1k vs 19.3k total tokens/solve) and approximately half the time (750 vs 1,481 seconds/case). This represents a substantial gain for users of the 3.8-27B model, as it reduces completion tokens from 12,547 to 7,301.

This benchmark uniquely details how UkisAI's Swift-Qwen3.8-27B model cuts reasoning tokens by 40% compared to the base Qwen3.8-27B, unlike other reports focusing solely on success rates.