Michael Limberger

Need me? Email mike@limberger.ca

← Skiff hub

I run a Mac mini M4 and a Mac Studio M2. The Mini is where Skiff and the test harness live. The Studio is my AI model server now: that is pretty much all it does. Local models run there with MLX, one at a time. Cloud models still go through Ollama on the Mini. Skiff does not call the vendor APIs directly. Keys stay in Ollama. This was for personal projects, not patient data.

Apple Silicon helps a lot here. CPU and GPU share one big pool of unified memory, so a large mixture-of-experts model does not have to shuttle weights back and forth over a PCIe hop the way a typical desktop GPU setup does. That is why these MoE packs feel practical on the Studio instead of stuck behind a smaller VRAM ceiling.

Hardware