Nativ

// Overview

Nativ is a macOS workspace for running AI models locally on Apple silicon, written in Swift by Prince Canuma, the author of mlx-vlm. It bundles an mlx-vlm server inside a SwiftUI app, finds compatible models already sitting in your Hugging Face cache (honouring HF_HUB_CACHE and HF_HOME), and puts chat, a model library, a performance dashboard and a local API server behind one window.

The chat side handles streaming conversations, image attachments, reasoning output and persistent history, with a separate Images tab for generating and editing pictures using compatible MLX image models. The model library browses and downloads models from Hugging Face with fit warnings sized against your machine's memory, and can preload separate language, image-generation, speech and embedding models at once — warning you when the combination would not fit.

The other half is operational. A performance page tracks request volume, token usage, time to first token and decode speed; a system monitor reports per-core CPU load, GPU utilisation, unified-memory and swap pressure, disk throughput and thermal sensors; and the menu bar can start or stop the server and pin live CPU, GPU and RAM graphs. Because the server speaks both OpenAI-compatible and Anthropic-compatible routes, the app doubles as the local backend for terminal coding agents and editors — the docs cover per-tool setup for Codex, Claude Code, OpenCode, Aider, Goose, Crush, Qwen Code, OpenClaw, VS Code, Cursor, Zed, JetBrains, Cline and Continue. MCP servers can be connected and their tools exposed to chat, and capabilities ship as independently versioned extensions, with a voice-dictation audio extension included.

// What it does

• Bundles an mlx-vlm server: no separate Python install to babysit, and models already in your Hugging Face cache are discovered automatically
• OpenAI-compatible chat, Responses, image, audio, embeddings and model endpoints plus Anthropic Messages endpoints, served on 127.0.0.1:8080 by default
• Model library with memory fit warnings, and simultaneous preloading of language, image, speech and embedding models
• Performance analytics (requests, tokens, time to first token, decode speed) and a live system monitor for CPU, GPU, unified memory, disk and thermals
• One-click configuration for terminal coding agents and editors to run against the locally served model
• MCP server connections and an extension platform, with a voice-dictation audio extension included

// Getting started

Nativ requires macOS 26 or newer on Apple silicon. Install the app from the project homepage or the GitHub releases page, then point your existing tools at its local server.

open https://github.com/Blaizzy/nativ/releases/latest

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "your-model-id",
    "messages": [{"role": "user", "content": "Why is the sky blue?"}],
    "stream": false
  }'

make xcode-generate
make xcode-build
make xcode-smoke

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Run private chat and vision models on a Mac with no data leaving the machine
• Serve a local OpenAI- or Anthropic-compatible endpoint to coding agents and editors you already use
• Compare MLX models by watching time to first token and decode speed while they run
• Watch unified-memory and thermal pressure while sizing which local models a Mac can actually hold

// How Nativ compares

Nativ alongside other open-source local runtimes tools AI/TLDR tracks, ranked by GitHub stars.