oMLX: guides, API, releases and benchmarks

Updated on 27 October 2026. Latest stable oMLX release: 0.7.2, from 21 October 2026.

oMLX is a local inference server for Apple Silicon Macs that serves MLX models through the OpenAI and Anthropic APIs, managed from the menu bar. This page puts what we have published about oMLX on jacar.es in order, from installation to the API and memory management. It also gathers the notes for each release, the comparisons with Ollama and LM Studio and our measurements on a 128 GB M5 Max.

oMLX in five facts

• Version: 0.7.2 is the latest stable release, published on 21 October 2026. 0.7.3rc1, from 27 October, is a release candidate under test.
• Requirements: macOS 15 or later and an M1 to M5 Apple Silicon chip. There is no Windows, Linux or Docker version.
• Address: the API listens on http://127.0.0.1:8000/v1 and the admin dashboard on http://127.0.0.1:8000/admin.
• API: compatible with OpenAI (/v1/chat/completions, /v1/responses) and with Anthropic’s Messages API (/v1/messages), so it serves Claude Code and Codex.
• Licence: Apache 2.0, in the jundot/omlx repository on GitHub.

Where to start

If you are new to oMLX, read the guides in this order. Each one solves one task and links to the next:

1. What is oMLX and how it differs from MLX, Ollama and LM Studio: what each layer is and when oMLX beats the alternatives.
2. How to install and tune oMLX on M5 Max 128 GB: the main guide, with installation, server settings, the model stack and the Claude Code connection.
3. Installing, updating and uninstalling oMLX with Homebrew: the command-line route and the service that starts on its own.
4. oMLX API reference: endpoints, API key, port and error codes: the reference to connect any client.
5. The oMLX admin dashboard and its command line: loading models, tuning each one and running tests.
6. Model management in oMLX: downloads, aliases, TTL and disk cache: which models stay in memory and which leave.

Release notes

Between 15 and 27 October 2026, oMLX published six releases across stable, candidate and development builds. These are the ones we have analysed or measured:

• 0.7.2 (21 October 2026): what changes in oMLX 0.7.2 and how to upgrade it on your Mac.
• 0.6.4 (29 August 2026): the advanced settings of 0.6.4 and its benchmarks on an M5 Max, inside the main guide.

The full list of changes in each release is in the GitHub release notes. Read them before upgrading a work Mac: a minor release can change the engine of a model you already use.

Comparisons

• oMLX against Ollama 0.40 and LM Studio on a Mac: engine, APIs, memory management and which to pick, from the October 2026 documentation.
• Ollama 0.40 runs MLX by default on Apple Silicon: what changed in Ollama and what it means if you already use oMLX.
• Local LLM calculator: whether a model fits in the memory oMLX lets you use on your Mac, and at roughly what speed.

Benchmarks and performance

• Local LLM benchmarks on Apple Silicon: all our oMLX measurements on a 128 GB M5 Max, with date, version and a downloadable CSV.
• oMLX benchmarks: measurement method and the memory ceiling: which variables to fix so a figure is comparable, and what to do when a model does not fit.

Integrations and uses

• Claude Code on oMLX: omlx launch claude and the variables it sets.
• oMLX as an MCP server: control your local models from Claude.
• Qwen and Clef on oMLX: one model writes and the other decides.
• Speech with oMLX TTS models: generate audio and clone a voice on your Mac.

Frequently asked questions

What is the latest oMLX version?

The latest stable release is 0.7.2, from 21 October 2026. 0.7.3rc1, published on 27 October, is a release candidate. The menu bar app announces each new release, and Homebrew installs it with brew upgrade omlx.

Is oMLX the same as MLX?

No. MLX is Apple’s compute framework, and oMLX is an independent server that wraps it and exposes it over HTTP. The difference is explained in what is oMLX.

Does oMLX have official documentation?

Yes: the README and the docs folder in the GitHub repository, and the omlx.ai website. The guides on this page are not official; they are checked against each version’s code and against a running server.

The Spanish version of this page is oMLX: guías, API, versiones y benchmarks.