Prince Canuma, the developer behind the widely used MLX-VLM library for running vision-language models on Apple Silicon, has released Nativ, a native macOS desktop application that wraps Apple's MLX framework in a full graphical interface.
The app functions much like LM Studio, offering both a chat window for direct interaction and a localhost API server that other applications and coding agents can call into. One detail early users have flagged as a thoughtful touch: Nativ automatically detects MLX models already sitting in a user's Hugging Face cache directory, so switching from a command-line workflow does not mean re-downloading everything.
Nativ runs open language, vision, video, code, and embedding models entirely on-device on Apple Silicon hardware. The project is free and fully open source, including the desktop client itself, and ships with a browsable model library covering compact vision-and-audio models as well as larger code-focused checkpoints. The pitch, according to the project's own materials, is a pointed contrast with other 'local AI' apps that are proprietary shells built over open-source inference engines they do not own. Nativ's team says every part of the stack, from the model loader to the telemetry charts, is visible and forkable.
The app is designed to plug into existing coding-agent workflows too: it exposes one local model endpoint that any compatible agent can be pointed at, without requiring an external API key once a model is downloaded.
The release picked up traction on Hacker News, gathering over 150 points and dozens of comments within a day of posting. One recurring observation from that thread is technical rather than promotional: Nativ's client is written primarily in Swift, which several commenters noted could make it comparatively straightforward to port the same inference stack to iPadOS and iOS in the future, since Apple's frameworks are shared across those platforms.
Nativ enters a local-inference space that already includes Ollama, LM Studio, Jan, and Locally AI, each competing on ease of setup, model breadth, and how tightly the tool integrates with a user's existing developer tools. Nativ's differentiator so far appears to be its depth of Apple-platform integration: it is built specifically around MLX rather than being a cross-platform wrapper, which should translate into tighter memory management on M-series chips at the cost of not running on Windows or Linux at all.
For now, inference itself runs fully offline once a model is downloaded, though first-time setup and model downloads still require network access. Support for dedicated audio-only and image-generation models is described as coming in a future update.
Whether Nativ becomes a serious alternative to LM Studio and Ollama on the Mac will likely hinge less on this initial reception and more on how quickly its open model library grows, and whether the promised mobile port materializes.