Sources
A documented 48GB M4 Pro Mac mini local-agent stack: 325 tok/s prefill, 34 tok/s generation, reachable over Tailscale
This 2026-09-01 writeup gives the numbers most local-LLM posts omit: Qwen3.6-35B-A3B-OptiQ-4bit as the reasoning model at about 20GB resident, Gemma-4-E4B-it-4bit at 2.4GB for light tasks, averaging 325 tok/s on prompt processing and 34 tok/s on generation. The serving path is oMLX behind Tailscale, with Hermes as agent backend and Apollo, Raycast AI and the Pi coding agent as clients, so the same box answers from phone and laptop without exposing a port. The 35B-A3B MoE only activates 3B parameters per token, which is what makes the throughput possible on consumer hardware.
Source
↳ Follow the thread