MLXHub: Local AI & LLM Server

Now runs across devices: pool memory from several iPhones and iPads on the same Wi-Fi to load a model none of them fits alone. Experimental, and honest about it.

Run AI locally. Your device. Your data.

MLXHub brings open-source language models to iPhone and iPad, powered by Apple Silicon and the mlx-swift inference engine. Your conversations never leave your device.

Distributed inference across your devices (Experimental) Pool memory from several Apple Silicon devices on the same Wi-Fi network to run a model that none of them could hold alone. One device hosts; the others join and each takes a slice of the model's layers — no device stores the whole thing. The host sends each device only the files its own layers need, and caches them so the next session starts faster.

What to expect: every device must be on the same Wi-Fi, you approve each device once by matching a code on both screens, and MLXHub measures each device's memory before it splits anything. This is an experimental feature — distributed generation works, but a run can still stall or drop a device, and not every model architecture can be split yet. It pays off most with large models. Joining a mesh is free on any device; hosting one is part of MLXHub Plus.

Chat with powerful models Download and run LLMs and vision-language models (VLMs) directly on your device. Send text messages or attach photos — the model sees and responds without touching any server.

Apple Intelligence built in On supported devices (iPhone 15 Pro / 16 and later with Apple Intelligence enabled), use Apple's on-device model for instant, private responses alongside any downloaded model.

Your personality, your assistant Set a global Agent Personality to define the AI's tone and style. Override it per-conversation with custom system instructions. No prompt engineering needed — just describe what you want.

Browse and install models Explore a curated catalog of optimized models — from tiny 0.6B models that fit in under 1 GB to powerful 7B models for deeper reasoning. Color-coded RAM badges tell you at a glance whether a model fits your device. Search HuggingFace directly to install any compatible model.

Local LAN server Turn your iPhone or iPad into a portable, chat API-compatible inference endpoint on your local network. MLXHub's optional LAN server exposes /v1/chat/completions so any app — from a Mac running OpenCode to a custom script — can use your device's models. No internet required. Auth-protected. Bonjour-discoverable. Part of MLXHub Plus.

Built for Apple Silicon MLXHub uses mlx-swift, Apple's own machine-learning framework, to run models at full Metal GPU speed. Model weights are quantized (4-bit, 8-bit) for the best quality-per-GB ratio on iPhone and iPad hardware.

Privacy Model inference happens on your device and never leaves it. No account. No ads. The optional LAN server only listens on your local network and is off by default. MLXHub collects a small amount of anonymous diagnostic data to find crashes and broken model loads; it is not linked to you, and you can turn it off in Settings.

Terms of Use (EULA): https://www.apple.com/legal/internet-services/itunes/dev/stdeula/