Source: Simon Willison's Weblog — 2026-07-21

Summary

Nativ, built by Prince Canuma (creator of the MLX-VLM project), is a free, open-source macOS app that runs open language, vision, and embedding models entirely on Apple Silicon and exposes them locally through an OpenAI- and Anthropic-compatible API. It's built on top of mlx-vlm, auto-detects models already cached from Hugging Face, recommends the right model size for your hardware, and does inference with no account, no subscription, and no data ever leaving the machine.

Key Takeaways

• Nativ wraps mlx-vlm — a package for inference and fine-tuning of vision-language models on Apple's MLX framework — in a native SwiftUI app, so it's a polished interface over an already-established open-source inference engine, not a from-scratch runtime.
• It exposes a local server with OpenAI- and Anthropic-compatible API endpoints, meaning existing tools and coding agents built against those APIs can point at localhost instead of a cloud provider with no code changes on the client side.
• It functions as four things in one app: a private chat client, a model manager that scans your Hugging Face cache for compatible models, a performance dashboard, and the local inference server itself.
• Support spans multimodal models (language, vision, and — via Canuma's related mlx-audio project — speech), tuned specifically for Apple's unified memory architecture and Metal acceleration rather than being a generic cross-platform port.
• Fully MIT-licensed and free, positioning it against hosted inference-as-a-service products on the specific dimensions of privacy (every token generated on-device) and zero recurring cost, at the tradeoff of being capped by whatever a given Mac's unified memory can hold.

Reel Script

Hook (~17s, 39 words) Every token you send to a cloud AI API leaves your machine, gets billed, and gets logged somewhere. A new open-source Mac app called Nativ runs real open models entirely on your own hardware — and speaks the exact same API your tools already expect.

Core Concept (~55s, 122 words) Running AI models "locally" usually means wrestling with command-line tools, picking the wrong quantization for your hardware, and hoping your setup doesn't silently swap to disk. Apple's MLX framework was built specifically to run models efficiently on Apple Silicon's unified memory, where the CPU and GPU share the same RAM instead of copying data back and forth like a traditional GPU setup — that's the "fingerprint" difference that makes Mac-native inference fast. Nativ takes mlx-vlm, an existing open-source engine for running vision-and-language models on MLX, and puts a real app around it: it finds models you've already downloaded, tells you which ones your specific Mac can actually run well, and starts serving them locally.

Hands-On (~50s, 110 words) The detail that matters for anyone building agents or tools: Nativ doesn't just chat, it serves an API shaped like OpenAI's and Anthropic's own endpoints. That means a coding agent or script written to call api.openai.com or api.anthropic.com can instead point at Nativ's local server address, and — for whatever open model you've loaded — nothing else about the client code has to change. Combine that with the model manager scanning your existing Hugging Face cache and a live performance dashboard showing what your hardware is actually doing, and you get a self-contained loop: pick a model sized for your Mac, serve it locally, point your existing tools at it.

Takeaway (~22s, 47 words) For prototyping, privacy-sensitive work, or just cutting your API bill, an OpenAI/Anthropic-compatible local server is the real unlock here — not the chat UI. If your tooling already talks to those APIs, swapping the endpoint to Nativ costs you nothing to try. Worth a look this week.

Discussion