Rapid-MLX is a local AI engine for Apple Silicon that provides a drop-in OpenAI-compatible API, with multiple model backends and support for real-time serving on macOS. It emphasizes speed and local execution, with various installation paths and a configurable RAM-based model selection.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Rapid-MLX is a local AI engine designed for Apple Silicon that offers a drop-in OpenAI/Anthropic API and enables running models locally on macOS. It advertises faster performance than Ollama and supports tool-calling, multiple model parsers, prompt caching, reasoning separation, and cloud routing. It includes an OpenAI-compatible HTTP server and CLI to launch, serve, or connect clients to local models. The README highlights the ability to serve via an HTTP endpoint and to wire up clients like Cursor, Aider, LangChain, and others to http://localhost:8000/v1 (and / for Anthropic endpoints).
How it works
The project provides a local inference engine with native Apple Silicon MLX kernels and a drop-in API, enabling chat, embeddings, audio, and video workflows. It supports a local server at http://localhost:8000, which clients can target using standard OpenAI/Anthropic SDKs. The Quick Start shows commands to install, run rapid-mlx chat, serve models, and launch IDE clients. It also describes optional extras for vision, audio, video, embeddings, and other capabilities as opt-in. The documentation indicates multiple models and tiers, with RAM-based selection guiding which alias to serve.
Getting started
Getting started steps (as shown in the Quick Start):
- Install: curl -fsSL https://rapidmlx.com/install.sh | bash or brew install rapid-mlx
- Chat with a model now: rapid-mlx chat
- Serve for use from apps: rapid-mlx serve qwen3.5-4b-4bit
- Wire up coding agent: rapid-mlx launch claude-code
The README provides commands for serving and for patching Claude Code to route at http://localhost:8000, plus an example OpenAI client configuration using base_url="http://localhost:8000/v1".
Recent releases
Latest releases include v0.12.4 (2026-08-04), v0.12.3 (2026-08-04), v0.12.1 (2026-08-03), and v0.11.9 (2026-08-02). Notable entries mention hardened benchmark methodology metadata and front-door tier map adjustments. The first signed + notarised rapid-mac release was v0.12.1 (2026-08-03) with a note about avoiding Gatekeeper warnings. The release notes show version bumps and fixes across these versions.
Traction
GitHub stars: 3400. The repository has 387 forks and 45 open issues. The project is written in Python and licensed under Apache-2.0. It was created on 2026-02-25 and last pushed on 2026-08-04. The project is categorized under topics including apple-silicon, fastapi, local-llm, m1, m2, macos, mlx, ollama-alternative, tool-calling.
Behind the repo
Rapid-MLX is associated with rapidmlx.com and the docs site rapidmlx.com/docs. The README references a model mirror at models.rapidmlx.com and a desktop app at rapidmlx.com/desktop.
Caveats
License: Apache-2.0. The README indicates multiple optional extras for vision, audio, video, and other capabilities that ship as opt-in. System requirements include Python 3.10+ as shown by the badges, and Apple Silicon support for M1/M2/M3/M4. The release notes mention that 0.12.0 internal build hit a Gatekeeper warning, with signed/notarised release in 0.12.1. The hardware-tiers section implies RAM-based serving configurations and requires sufficient memory for larger models.