Nativ, an open-source macOS app, runs frontier open AI models locally on Apple Silicon

Nativ, a free, open-source macOS app that runs open AI models locally on Apple Silicon, was released July 20, 2026 by the developer Blaizzy, who also created the MLX-VLM toolkit, according to the project site.

The MIT-licensed app runs language, vision, video, code and audio models on Macs with Apple’s M1 chip or later through the MLX-VLM framework, with no accounts, subscriptions or cloud dependency. It bundles a local inference server, auto-detects compatible models from a user’s Hugging Face cache, and shows live telemetry including tokens per second, memory use and thermal state.

The pitch is one-click local inference for developers who want to keep data off the cloud. Nativ integrates with coding tools including Claude Code, Codex and OpenCode, letting those tools call a locally hosted model instead of a remote API. The source code is on GitHub.

Local inference apps have multiplied as open models close the gap with hosted ones, and Apple Silicon’s unified memory makes consumer Macs viable for models that once needed a server. Nativ’s angle is packaging that into a single app rather than a command-line setup.

The project is the work of one developer, and its capabilities rest on that developer’s own description; there is no independent testing of the models’ speed or output quality across the range Nativ claims to support.

Whether Nativ gathers a contributor community and independent benchmarks, or stays a solo side project, will determine if it becomes a fixture of the local-inference toolkit.