MLX-Serve
Run open models on your Mac: chat, coding agents, images, video, music, speech and 3D. One native app and one local server that speaks the OpenAI, Anthropic and Ollama APIs. It runs offline, needs no Python and no account, and it's MIT licensed.
Download for Mac · Homebrew · iPhone · Linux · mlxserve.com · 简体中文
Install
The app. Download MLX-Serve, drag it to Applications and open it. It needs macOS 26.2 or later on Apple Silicon, suggests a model that fits your Mac, and updates itself.
Homebrew.
brew tap ddalcu/mlx-serve https://github.com/ddalcu/mlx-serve
brew trust ddalcu/mlx-serve
brew install --cask mlx-serve # the app
brew install mlx-serve # or only the command-line server
The terminal, if you know Ollama:
mlx-serve run gemma4 # downloads Gemma 4 E4B, serves it and chats right here
mlx-serve pull gemma4:26b # download only, resumable
mlx-serve serve # serve everything you downloaded; models load when asked for
The server listens on http://localhost:11234. Point any OpenAI, Anthropic or Ollama client at it, or open that address in a browser for the built-in web console.
iPhone
MLX-Serve puts the same engine on your iPhone: chat with Gemma 4, clone your voice, and make pictures and music, all on the phone and offline. It needs an A17 Pro-class iPhone or newer with iOS 26.
Linux and Windows
Every release has Linux downloads for NVIDIA GPUs: x64 for RTX 20 to 50 cards, arm64 for DGX Spark, GH200 and Jetson Thor. On Windows with an NVIDIA card, one script builds it. Details are in docs/building.md.
What it does
- Coding agents on a local model.
mlx-serve launch claude starts Claude Code against your Mac, set up with the model's real context window. Codex, OpenCode, pi, Grok and others work the same way. Integrations
- Pictures, video, music, speech and 3D. FLUX.2, Qwen-Image, Krea, MOSS-Image, LTX-Video, MiniMax-H3, ACE-Step, YuE2, Qwen3-TTS with voice cloning, and Hunyuan3D, in the app or over HTTP. The app
- Transcripts and meeting notes. Speech to text in 30 languages, with word timings and who said what. The menu bar records a call and writes the notes. API
- Models bigger than your Mac. GLM-5.3-Flash and DeepSeek-V4.1-Flash run on a 128 GB Mac by reading their experts from the SSD. Two Macs on a Thunderbolt 5 cable can run one model together, and Macs on your network can use each other's models. CLI · FAQ
- An assistant that lives on your Mac. Chat with your documents, talk to it hands-free, give agents a sandboxed shell and MCP tools, or schedule tasks in plain English. The app
- The APIs your tools already use. OpenAI chat, completions and Responses, Anthropic Messages and Ollama, all on one port, with streaming, tool calls, JSON schemas, vision and reasoning. API
What can my Mac run?
The app suggests a first model for your Mac's memory:
| Memory | Suggested model |
|---|
| 16 GB | Gemma 4 E4B |
| 18–24 GB | Gemma 4 12B |
| 32 GB | Qwen 3.8 27B |
| 36 GB | Qwen 3.8 27B 6-bit |
| 48–64 GB | Qwen 3.8 27B 8-bit |
| 96 GB and up | Qwen3.8-Flash-Next (125B, about 6B active per token) |
| 128 GB | Also GLM-5.3-Flash and DeepSeek-V4.1-Flash, their experts streamed from the SSD |
| 256 GB | Also GLM-5.3-Flash in memory (NVIDIA's FP4 pack) |
Thousands more run too: any MLX model with a supported architecture, and any GGUF. The tier list ranks the popular ones, and docs/models.md lists every architecture.
Speed
Decode speed in v26.11.1 on an M5 Ultra with 256 GB, on a code prompt, with speculative decoding on:
| Model | Tokens per second |
|---|
| Qwen3.6 35B-A3B, 4-bit | 421 |
| Qwen3.8 27B, 4-bit | 243 |
| Gemma 4 E4B, 4-bit | 200 |
| Gemma 4 26B-A4B, 4-bit | 184 |
| Qwen3.8-Flash-Next, 4.7 bits | 180 |
On an M4 Max, the same models decode 10% to 51% faster than they did in v26.8. Every release is measured the same way in benchmarks.md; speculative decoding and tuning are in docs/performance.md.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
