vMLX

About

vMLX is a Mac app for running AI models entirely on Apple Silicon, with no API keys, no subscription, and nothing uploaded to the cloud. It lets you chat with open models, write code using 20 built-in agentic tools (file I/O, search, shell, git, and web access) with native MCP support, and generate or edit images with nine local models, including Flux and Qwen Image variants. It can also convert models from GGUF to MLX inside the app and serve OpenAI- and Anthropic-compatible APIs to other applications such as Claude Code. The app is free and requires macOS 14.5 or later, M1 through M5 chips, and at least 8 GB of memory (16 GB or more recommended).

MLX Studio is the app itself, covering the chat window, tools, and model browser. Underneath it runs vMLX, an open-source inference engine that includes prefix caching, paged KV cache, cache quantisation, continuous batching of up to 256 concurrent sequences, and a persistent disk cache. The website reports benchmarks against LM Studio on an M3 Ultra with 256 GB running Llama 3.2 3B at 4-bit, showing time-to-first-token improvements from 9.7× on a 2.5K-token prompt to 224× on cold processing of a 100K-token prompt. For users who only want the server, vMLX is also available as a pip package.