Training Qwen 4B to Beat Large Models on Work Tasks