Nativ, and the quiet case for keeping the model on your Mac

Prince Canuma, who wrote the excellent MLX-VLM library, has wrapped MLX in a real macOS app called Nativ — a chat interface, a model manager, a performance dashboard, and a localhost server that speaks the OpenAI and Anthropic APIs, all running on Apple Silicon. Simon Willison notes it picked up the MLX models already sitting in his Hugging Face cache, which is the kind of small courtesy that tells you someone was paying attention to how people actually work.

What we like about it isn’t the feature list; it’s where the computation happens. A local inference server means the prompts, the documents, the half-formed ideas you feed a model never leave the machine. Privacy stops being a policy you have to trust and becomes a property of the architecture — there’s no server to leak, no telemetry to opt out of, no terms of service that quietly change. The unified memory that makes M-series Macs good at this is the same thing that makes the private version practical.

This is a principle we keep coming back to as we build: the most respectful software treats your machine as yours, and does its work where you can see it. A native app that runs frontier-ish models on your own hardware is a small proof that the private option doesn’t have to be the worse one.