Pre-loaded, not locked down
Models are already on it. They're not the deal.
Your mini ships with a working local stack — a runtime, an OpenAI-compatible endpoint on the loopback, and a set of open-weight models sized to fit 64GB. Plug in and you're generating tokens the same day, without spending a weekend on build flags and quantization math.
That's a starting point, not a product boundary. You get root over SSH. Pull anything you want from Hugging Face, swap llama.cpp for MLX or vLLM, run whatever quantization you prefer, or delete our entire stack and start from an empty disk. Nothing we install is required to keep the machine running, and nothing phones home about what you run.
Open box, not a managed API.
There's no allowlist of approved models, no per-token metering, and no proxy in front of your endpoint. The weights on the disk are yours to change, and the 4TB gives you room to keep a dozen of them on hand instead of re-downloading every time you switch.