Jeff-Code: a 0.8B model makes Qwen 3.8-27B coding 47% faster, same pass rategithub.com·2 pts·firelex·0