Nativ: Run AI models locally on your Mac
Simon Willison Simon Willison
Prince Canuma, the guy behind MLX-VLM, just shipped Nativ, a native macOS app for running AI models locally with MLX. It's basically a slicker, homegrown LM Studio alternative for Mac users who want offline models without touching Python.
Prince Canuma has quietly been one of the more useful figures in the local-AI-on-Mac scene, building MLX-VLM to let people run vision-language models through Apple's MLX framework. Now he's taken that work and wrapped it in something far more approachable: a full desktop app called Nativ, built for macOS, no terminal required.
The pitch is straightforward. Nativ gives you a chat interface for talking to models running entirely on your own machine, plus a localhost API server so other apps and scripts can hit those same models over HTTP. That's the same basic shape as LM Studio, which has become something of a default choice for local model runners on Mac. Nativ isn't trying to reinvent that formula, it's trying to do it with MLX under the hood, which matters if you care about squeezing performance out of Apple Silicon specifically rather than relying on more generic backends.
What stood out, based on early hands-on impressions from developer and blogger Simon Willison, is a small but telling detail: Nativ automatically found MLX models that were already sitting in his Hugging Face cache directory. No re-downloading, no manual pointing to file paths. It just picked them up. That's the kind of quality-of-life touch that separates a tool built by someone who actually uses this stuff daily from one built purely as a demo.
Canuma's background gives Nativ some credibility that a brand-new entrant wouldn't have. MLX-VLM has already proven itself as a solid way to run vision models locally, and folding that expertise into a polished GUI is a natural next step rather than a pivot. For Mac owners who've been curious about running open models offline but didn't want to fuss with Python environments, Nativ looks like one more reason to just try it.
My take
I like seeing more MLX-native tools instead of everything defaulting to generic backends that treat Apple Silicon as an afterthought — that's the whole appeal of local models, squeezing real performance out of hardware you already own. Nativ picking up my Hugging Face cache automatically is exactly the kind of detail that separates a tool made by a practitioner from one made for a demo video, and honestly the local-model tooling space needs more of that and less hype-driven vaporware.
Read more about this at: Simon Willison
