Qwen3.8-27B that thinks 58% less with near-equal accuracy
Qwen3.8-Flash-Next that thinks 63% less, within 1 point
Swift Bonsai 2, a 2-bit Qwen-based 27B reasoning model