Came across this popular open source inference engine Tensorfold https://github.com/ashhart/TensorFold Using their official Vontra/Qwen3.8-27B-MLX-4bit with drafting model z-lab/Qwen3.8-27B-DFlash2, I can reach 40-60 tps on mac mini m5 pro. AGAIN, it is PRO on mac mini, not even ultra studio. For me
