Qwen3.5-397B-A17B runs on 48GB MacBook Pro at 4.4 tok/s via pure C and Metal

How this story unfolded

5 days · 0 reports · 3 community posts · from Aug 14

1. Aug 14
   • Runs LLM inference without PyTorch, Triton, or transformers, compiling kernels directly to GPU code...
2. Aug 17
   • Runs the 397B parameter Qwen3.5-397B-A17B model on a MacBook Pro with 48GB RAM at 4.4+...
3. Aug 19
   • Runs frontier MoE models up to 2.8T parameters on consumer hardware using pure C with zero...