MLX Serve
ddalcu/mlx-serveReviewed by AI
Checked by RepoPicks Radar on
What is MLX Serve?
A Mac app plus a local server that runs open AI models on your own machine. You can chat with a model, generate pictures, video, music, speech and 3D, transcribe calls and write meeting notes, and point coding tools like Claude Code or Codex at the local server instead of a paid online service. The Mac app has an installer and suggests a first model for your amount of memory; it needs macOS 26.2 or later on Apple Silicon. Linux and Windows downloads exist but are command-line only and expect an NVIDIA graphics card.
MLX Serve is free and open source (Apache-2.0) and can replace ChatGPT, Midjourney, and OpenRouter.
How easy is it to install MLX Serve?
B · installer plus one simple step · macOS
- Windows D
- macOS B
- Linux D
On a Mac you download MLX-Serve.dmg, drag it to Applications and open it, and the app then suggests and downloads a model that fits your memory, which is the one extra step that makes it a B; it needs macOS 26.2 or later on Apple Silicon. Linux and Windows are harder (D): the downloads are command-line servers built for NVIDIA cards (CUDA 13), and the README says Windows needs a build script.
MLX Serve on GitHub
- Stars
- 2,420on Nov 10, 2026
- Downloads
- 68,3831,506 this week
- Growth
- +53stars in 6 days
Stars over time
Facts
- Runs on
- macOS, Windows, Linux
- You need to install
- Nothing extra
- Disk space
- 228 MB–683 MB · from the installer
- Accounts or keys
- None
- Cost
- Free · The README says nothing is paywalled and asks only for optional support through GitHub Sponsors or Buy Me a Coffee. The separate iPhone version is listed on the App Store.
- License
- Apache-2.0 (Permissive)
- Can replace
- ChatGPT Plus, Midjourney, ElevenLabs, OpenRouter, Otter.ai, Veo 3
- Last change on GitHub
- Nov 10, 2026
MLX Serve risk signals (4)
- Infolicense file mixes mit and apache
The LICENSE file starts with the MIT text ("Copyright (c) 2026 David Dalcu") and then says "This distribution also includes third-party software under its own licenses, including the Apache License 2.0, the BSD 3-Clause License, and further MIT-licensed components". That mix is why licence detectors disagree (GitHub reports no single licence for the repository); all of them are permissive.
AI review
- Infoapp updates itself
The install section says the app "suggests a model that fits your Mac, and updates itself", so after installing it will fetch new code on its own without going through the GitHub releases page.
AI review
- Infolarge model downloads at run time
The app and the commands download models when asked ("mlx-serve run gemma4 — downloads Gemma 4 E4B"), and the suggestion table goes up to models of 125B and 744B parameters "streamed from the SSD". Real disk use depends entirely on which models you pull, far beyond the 228 MB to 683 MB of the installers.
AI review
- Infohomebrew tap asks for trust
The Homebrew path is "brew tap ddalcu/mlx-serve https://github.com/ddalcu/mlx-serve" followed by "brew trust ddalcu/mlx-serve", which tells Homebrew to run code from this author's own tap instead of the official one.
AI review
Evidence is shown in its original language.
What Radar could not check about MLX Serve
- Whether it installs and runs on each Mac, and whether the macOS 26.2 and Apple Silicon requirement rules out your machine.
- The speed table in the README, measured by the author on an M5 Ultra with 256 GB.
- Whether very large models really work by reading from the SSD, and how usable they feel.
- Total disk space and memory after downloading the models you want.
- What the self-update and the app contact over the network, since the README says it can run offline but does not list the addresses.
- Whether the Windows and Linux builds work out of the box on a given NVIDIA card.