On device · iPhone & Mac
A private AI assistant for iPhone.
Language models run on your iPhone — or on your own Mac, when you want a bigger one. There is no account to create, and no server of ours in between.
Coming soon to the App Store
No account
Nothing to sign up for
No servers
We operate none
No analytics
Not even a crash SDK
What can leave your phone
Two features, both off until you turn them on.
Everything else — generating replies, transcribing your voice, and reading your documents — happens on the device.
Web search
Sends a derived query — never the conversation — to a provider you supply your own key for. Email addresses, phone numbers and long digit strings are stripped first.
Mac offload
Sends your prompt to a Mac you paired, so it can answer with a model far larger than a phone can hold. That Mac reads it in order to answer — which is not the same as never leaving, so the app says so.
What it does
Your own documents
Add PDFs or text files. Read, chunked and indexed on the device, with answers that cite the page they came from. Scanned pages are recognised on the device too.
Voice, both ways
Dictation transcribed on device; replies spoken back as they generate. Recordings are deleted the moment they are transcribed.
Several models
Apple's built-in system model, or anything in MLX format — from something small enough for a phone to something only your Mac can hold.
Threads that persist
Token streaming, Markdown, and context-window management, so a long conversation does not quietly lose its beginning.
Locked
Face ID on launch and a privacy cover in the app switcher. Keys never sync and never ride a backup to another device.
No signal needed
Once its models are downloaded, everything above works in aeroplane mode: on a flight, underground, anywhere.
Your Mac's model, from your phone. With nobody in the middle.
Pair the two once, with a code shown on the Mac and typed on the phone. The Mac's models then appear in the phone's model picker.
01
No account
Nothing to sign up for, and no list of your devices held anywhere. The pairing code is the whole relationship.
02
Forward secret
Fresh keys per connection, held only in memory. Traffic recorded today stays unreadable to someone who takes the pairing code tomorrow.
03
Post-quantum
A hybrid of X25519 and ML-KEM-768. It holds if either half falls — including to a quantum computer.
What fits
Pick your iPhone. See what runs on it.
Every model in the catalogue, drawn to scale. The built-in one needs no download at all.
6 of 8 models run on this iPhone.
- Apple Intelligencebuilt inBuilt in, starts instantly
- Llama 3.2 1B0.7 GBFastest; runs on any supported iPhone
- Qwen 3 1.7B1 GBNoticeably stronger than Llama 1B for the size
- Llama 3.2 3B1.8 GBSolid general purpose
- Qwen 3 4BPick2.3 GBBest all-round that still fits a 6 GB iPhone
- Gemma 3n E2B2.5 GB~5B parameters behaving like a 2B
- Phi 3.5 Mini2.2 GBNeeds 8 GB
- Gemma 3n E4B3.9 GBNeeds 8 GB
Download size is the weights. A long conversation also needs room for its cache, which is what actually ends one — why that matters.
What you need
iPhone. iOS 26 or later. Running a model on the phone wants recent hardware and room to spare; the app tells you before it downloads anything.
Mac (optional). macOS 26 or later, Apple silicon. Only needed if you want to offload to a larger model.