A 48GB M4 Pro Mac runs a local AI stack over Tailscale

A 48GB M4 Pro Mac is running two local models at once. The setup, detailed on lws.io, pairs Qwen3.6-35B-A3B-OptiQ at 4-bit quantization, about 20GB resident, for reasoning tasks with Gemma-4-E4B-it at 4-bit, 2.4GB, for lighter work. It serves both to a phone and a laptop over Tailscale, with no public port opened.

The setup gets 325 tokens a second on prefill and 34 tokens a second on generation, on a consumer machine rather than a data center card. That's fast enough to use as a daily driver instead of a demo you show someone once and close.

The architecture is what makes those numbers possible. Qwen3.6-35B-A3B is a mixture-of-experts model. It has 35 billion parameters total, but only 3 billion activate for any given token, which keeps memory and compute in reach of a Mac instead of a rack of GPUs.

The stack routes through oMLX behind Tailscale, with Hermes running as the agent backend and Apollo and Raycast AI as the clients people actually type into. Same box, same models, whether someone's on a phone on the couch or a laptop at a desk.

What the writeup doesn't cover is cost or setup time. There's no line on how long the quantization and configuration took, or what the M4 Pro Mac cost going in. Anyone weighing a similar build has to fill in that gap before copying the parts list.