Mac Mini
The always-on on-ramp: a silent, low-power desktop you can leave serving 24/7. Apple Silicon's unified memory lets a $600–2000 box hold models that would otherwise need a discrete GPU with the same VRAM — the deciding advantage for this tier.
What fits
| Config | Usable for models | Good picks | Tier |
|---|---|---|---|
| M4 Pro, 64GB unified | ~48GB | Qwen3 32B, Qwen3 30B-A3B (MoE — fastest big model here), Gemma 3 27B | executor / executor-heavy |
| M4, 16–32GB unified | 10–24GB | Qwen3 8B/14B, Gemma 3 12B | swarm / executor |
Budget ~25% of total RAM for macOS (a 64GB Mac → ~48GB for models). python scripts/fit_device.py --memory 48 --backend metal picks the exact model, context, launch command, and endpoints.yaml stanza.
Serving
llama.cpp (Metal) or MLX, via hardware/personal-devices/configs/serve-apple-silicon.sh:
MODEL=Qwen/Qwen3-30B-A3B-GGUF ./configs/serve-apple-silicon.sh # Metal (default)
MODEL=mlx-community/Qwen3-30B-A3B-4bit BACKEND=mlx ./configs/serve-apple-silicon.sh
Register under tier executor in scripts/endpoints.yaml.
Role
The cheapest way to stop paying frontier credits for routine execution — and the Mac Mini is built for exactly this: near-silent, sips power, happy as an always-on endpoint. Start here before investing in a larger swarm or an H100 node.
Desk-side execution is why Savings can look so large — please consider donating to help support this project.