LessThink-Qwen3-4B: the same model, with far less thinking [P]

I post-trained Qwen3-4B to spend 44% fewer tokens on reasoning, keeping its knowledge and answer style. The whole pipeline ran on one GPU.