Nativ Launches Open-Source MLX App for Local Mac Models
Nativ released version 0.0.1 on July 20 as an MIT-licensed macOS app for running MLX models locally on Apple silicon. It combines chat, model management, performance telemetry, and OpenAI- and Anthropic-compatible local APIs; independent developer Simon Willison reported that the app recognized MLX models already stored in his Hugging Face cache.
Nativ released version 0.0.1 on July 20, offering a native macOS workspace for running and serving MLX-format models on Apple silicon. The project is MIT licensed and provides a downloadable disk image alongside its public source code.
What shipped
Nativ combines a SwiftUI desktop interface with a bundled mlx-vlm server. The app can discover compatible models in a user's existing Hugging Face cache, download and switch models, run text and vision chats, and display metrics such as token usage, time to first token, decode speed, memory pressure, and recent request activity.
It also exposes local endpoints compatible with OpenAI chat, Responses, model, image, and audio routes, plus Anthropic Messages and token-counting routes. The project lists integrations for Codex, Claude Code, Pi, Hermes, and OpenCode, allowing those clients to send requests to a model served on the Mac.
The first release requires Apple silicon, macOS 26 or newer, and enough unified memory for the selected model. Model downloads and the first build from source still require network access; the inference workload runs locally after the model is available. The repository says dedicated audio-only and image-generation-only model support is planned rather than included in the current release.
Independent check and limits
Developer Simon Willison reported that Nativ recognized MLX models already present in his Hugging Face cache, a practical confirmation of one of the launch claims. His post described the app as similar in shape to LM Studio, with both a chat interface and a localhost API server.
This is an early v0.0.1 release, so its long-term stability, model compatibility, and performance across different Macs are not yet established by broad testing. Claims about privacy and local execution come primarily from the project; users should still review the open-source code, release artifacts, network behavior, and each downloaded model's license before adopting it for sensitive work.
For data and ML practitioners, the useful distinction is workflow consolidation. Nativ does not introduce a new model architecture. It packages MLX inference, model discovery, monitoring, and familiar API shapes into one native app, reducing the setup needed to test local models with existing agent and developer clients.
Key Points
- 1Nativ v0.0.1 packages an mlx-vlm server, chat interface, model manager, and performance dashboard in an MIT-licensed native macOS application.
- 2The app exposes OpenAI- and Anthropic-compatible localhost APIs and connects supported coding clients to models running on Apple silicon.
- 3The first release requires macOS 26 or newer and sufficient unified memory, while broader stability and compatibility remain to be established beyond early use.
Scoring Rationale
The release provides a useful open-source desktop and local-API layer for MLX workflows, with an independently observed cache-discovery feature. Its scope is practitioner-relevant but narrower than a new model or infrastructure breakthrough, and v0.0.1 has limited independent testing.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
