Optimized with MLX
Native inference built on mlx-swift, Apple's open-source machine learning framework, tuned for the GPU of every Apple Silicon chip.
opensource.apple.com/projects/mlxOn-device AI for iPhone & iPad
MLXHub
Your personal intelligence, wherever you go.
Run powerful language and vision models entirely on your Apple Silicon device.
Requires iOS or iPadOS 26
- No cloud AI
- No account
- Works offline
Built for Apple Silicon
Everything runs on your device.
Private by design
Your conversations stay on your device unless you ask for something online, like a web search. No account needed, and usage analytics never include what you write; you can turn them off in Settings.
No internet connection needed
Once a model is downloaded, it runs fully offline — on a plane, underground, anywhere.
Hugging Face downloads
A curated catalog of open models. Free and resumable.
Model mirrors
Switch to ModelScope or your own mirror when Hugging Face is blocked, and keep downloading.
Multiple model selector
Switch between text and vision-language models in a single tap, mid-conversation.
KV cache
Faster multi-turn chats with key-value caching.
Voice
Just talk. It listens, thinks and answers out loud.
Ask a question while you cook, rehearse a hard conversation on a walk, or think out loud at the end of the day. Your voice is transcribed and answered on your device, so what you say stays with you.
Voice mode in every chat
Tap the wave next to the message box and keep talking. Interrupt whenever you like: it stops, listens and carries on from there.
Voice Center on iPad
With the LAN server on (MLXHub Plus), prop an iPad in the kitchen or the living room and simply ask. Hold to ask once, or lock it for a hands-free conversation. It keeps no history.
Meet the Entity
It shows you what the model is doing and takes on the colour of the model family you choose.
Image creation
Describe it. Watch it appear.
Turn an idea into a picture: a card for a birthday, a mood for a project, the place you're trying to describe. Images are made on your iPhone or iPad, so your prompts and pictures stay yours.
- Create from the Images tab, or just ask in any chat.
- 512 px for everyone, up to 1280 px with MLXHub Plus.
- Every image is tagged as AI-generated in its file metadata.
Tools & Skills
Answers that know your day.
Ask how you slept this week, what's left on today's calendar or where to find a quiet café nearby. MLXHub only uses the apps you allow, and works out the answer on your device.
Health
“How did I sleep this week?”
You averaged 7 h 12 min a night, 25 minutes more than last week.
Reads your steps, heart rate and sleep from the Health app, and logs new data when you ask.
Calendar
“What's left for today?”
Two things: a call with Ana at 4:30 and the dentist at 6.
Checks your events and reminders, and adds new ones from a simple sentence.
Location
“Find a quiet café near me.”
Café Lento is a four-minute walk away and usually quiet in the afternoon.
Searches places around you and checks the weather where you are.
Also works with
- Reminders
- Apple Music
- Photos
- Weather
- Wallet
- Translate
Skills
With MLXHub Plus, teach it how you like things done once, and use it in any chat.
Example answers are illustrative. You decide which apps it can use. Tools that create or change your data are part of MLXHub Plus.
Bigger models
Run models bigger than your device.
Some of the best open models need more memory than any phone has. MLXHub finds a way: it shares the work between your devices, or reads what it needs from storage as it goes.
Your devices, working as one
Pair your iPhone and iPads, and a model too big for any one of them runs across all of them. At home over your Wi-Fi: paired, encrypted and on a network you control. Joining is free; hosting a model needs MLXHub Plus.
More model than memory
Models made of many specialised experts keep only what they need in memory and read the rest from storage while they answer. That is how a 35-billion-parameter model runs on an iPhone with 12 GB.
Qwen3.6 35B-A3B (20 GB) on iPhone 17 Pro Max with 12 GB of memory.
LAN server
Your iPhone becomes your home's AI server.
With MLXHub Plus, keep MLXHub open on the charger and it becomes a private AI for every device at home. It speaks the same language as OpenAI and Ollama, so the apps you already use simply work.
Includes a web chat for any browser on your network.
http://your-iphone.local:8080Running
OpenAI-compatible
- POST
/v1/chat/completionsChat - POST
/v1/completionsText completion - POST
/v1/embeddingsSearch and memory - POST
/v1/rerankRank results - POST
/v1/images/generationsCreate images - GET
/v1/modelsInstalled models
Ollama-compatible
- POST
/api/chatChat - POST
/api/generateGenerate - POST
/api/embedEmbeddings - GET
/api/tagsInstalled models
Search by meaning
Turn your notes, documents and code into a private search engine. Embeddings let your apps find things by what they mean, not just the words they contain, and can stay loaded alongside the chat model when memory allows.
A voice for the whole home
Turn on the Voice Center on an iPad and anyone at home can simply ask out loud. The answer is worked out on your own devices, never in a cloud.
Your tools, one address away
Code editors, notes apps and your own scripts connect by changing a single URL. No accounts, no API keys from a cloud provider.
Runs popular open models, RAM-aware for your device
Explore the full catalog on Hugging FaceQwen3.6 35B A3B
Qwen
Bonsai 27B
Prism ML
LFM2.5 8B A1B
Liquid AI
Gemma 4 E4B
Google
Gemma 4 E2B
Google
Frequently asked
Which devices are supported?
Any iPhone or iPad running iOS or iPadOS 26 or later. Which models fit depends on your device's memory, and MLXHub checks before you download.
Does it cost anything to run models?
Downloading and running models is free, with no per-token or API fees. MLXHub Plus adds the LAN server, hosting a model across your devices, Skills, tools that change your data and larger images.
Is my data really private?
Chats run on your device and are never sent to a cloud AI. Online features you turn on, like web search, downloads, the LAN server or pairing devices, talk only to their service or to your own devices. There is no account and no ads. Analytics and diagnostics never include what you write, and you can turn them off in Settings.
How does the LAN server work?
With MLXHub Plus, your iPhone or iPad serves OpenAI- and Ollama-compatible routes (chat, completions, embeddings, rerank, images and models) and a web chat on your local network, on port 8080 by default. It only answers connections from your local network and runs while MLXHub is open in the foreground.
Will it work offline?
Completely. After a model is downloaded, everything runs without an internet connection.
How do I know a model will fit?
We do our best to indicate this right in the app, though the metadata Hugging Face provides is sometimes inconsistent — so a model that doesn't actually fit can occasionally still be downloaded. When that happens, we try to stop it from loading into RAM before iOS triggers a jetsam (an abrupt app shutdown), but we can't always guarantee it.
Why do bigger models fit on the iPad than on the iPhone?
Apple enforces stricter memory limits for apps on iPhone than on iPad, mainly to protect system stability on a smaller device. We do everything we can to make the most of the RAM headroom iOS grants us, but some of those ceilings are hard limits that only Apple can change over time, typically as new devices and OS versions raise them.