Prince Canuma, developer of the MLX-VLM Python library for running vision-LLM models on Mac with Apple’s MLX framework, has released Nativ. It’s a macOS desktop app that wraps MLX in a complete interface: chat and localhost API server, resembling LM Studio. The app recognizes MLX models already present in your local Hugging Face cache, so you don’t redownload them.

Why it matters. If you have a Mac and want a model to read images without sending your data to the cloud, until now you had to configure MLX-VLM from the terminal. LM Studio and Ollama cover text models locally well, but vision-LLMs (models that combine text and images) remained the domain of Python developers. Nativ brings local multimodal to those who prefer a GUI. It’s part of a trend we’re tracking: local inference is spreading across different form factors, from 1-bit models in the browser to LM Studio Bionic, and now to vision-LLMs.

If you want to try it. Starting point: download the app from the project site and point it to an MLX model already downloaded from Hugging Face.

In detail

MLX is Apple’s open source framework for machine learning on Apple Silicon, designed to leverage the unified memory of M chips. Unlike CUDA, which requires dedicated NVIDIA GPUs, MLX runs on any Mac with M1 or later. MLX-VLM, also by Canuma, is the next layer up: it takes vision-language models (architectures like LLaVA, Idefics, Qwen-VL) and adapts them to run with MLX. Until now, using it meant installing Python, managing dependencies, and running scripts from the terminal.

Nativ closes that gap. The app detects models already downloaded in your Hugging Face cache folder, lists them in a chat interface, and also offers an API endpoint on localhost, following the same logic as LM Studio. You can point another tool or script to the same port and use the model as if it were a cloud API, except it runs on your Mac and nothing leaves the machine.

What differs from LM Studio: LM Studio primarily supports text models in GGUF format via llama.cpp, and has added some vision capabilities incrementally. Nativ starts from the opposite assumption: it builds on Apple’s native MLX and targets vision-LLMs first. The two tools can coexist: those who already have models cached will find them in both.

The most concrete use case is local multimodal. You have a contract scan, a screenshot of an interface, a product photo: you give it to the model and ask it to extract data, describe what it sees, summarize. Until today, doing this without the cloud required Python and patience. Now you just drag the file into the chat.

Real limitations. The app is new and Mac-only: no Windows, no Linux. The availability of vision models in MLX format depends on who converts them, and not all models have a ready MLX equivalent. Inference quality on Apple Silicon holds up for small and medium models, while frontier models remain out of reach for memory and speed. Finally, it’s a solo developer project: long-term support depends on how much the community builds on it.

For those working with visual documents or sensitive data and wanting to experiment with vision-LLMs without the cloud, it’s the first serious attempt to make this accessible to non-programmers.