Own AI
06 Turns
Only what is new gets read.
Each turn the whole chat is rendered exactly as the model was trained on it. The part the cache already holds is skipped, and only the rest is fed in.
Own AI · on-device assistant for iPhone
Private AI, at home on your iPhone.
Own AI runs language and vision models on the device itself. Chat, ask your documents, talk hands-free and work through math, with a starter model already inside.
- 88downloadable MLX models
- 13model families
- 6local voices
- 4languages: EN · DE · ES · FR
Fig. 01 Where the thinking happens
With a local model, your question, documents, photos and voice are processed on the iPhone
Elsewhere, asking means sending.
Here, the model is already on the phone.
So a question, a lease, a photo and your voice are read where they are.
Local inference sends no prompt to a model provider.
Stored on the device
Conversation history, imported documents, retrieval indexes, saved prompts and assistant memory.
Declared in the privacy manifest
No tracking, and no collected data.
Only when you ask for it
Model downloads, web search and Apple Intelligence talk to their own services, and only when you use them.
Fig. 02 Ask your documents
Hand it the lease. It points at the page.
Attach PDFs, DOCX, TXT, RTF or code files and ask. Retrieval runs on the phone, scans can go through optional OCR, and Document Sources shows the pages an answer drew on.
- Grounded answers, with their sources on show.
- Searched entirely on your device.
- Keep going in the same chat, like drafting the email to the landlord from what the lease actually says.
A moving sketch of the flow. The question and answer are from the app's sample lease.
- Document Sources
- Asking a lease about its fees
- A follow-up from the same file
Fig. 03 Math that draws itself
Equations unfold, then plot.
Answers render LaTeX and group a derivation into numbered steps. Any formula opens in the Formula Inspector, with four tabs: Formula, Steps, 2D and 3D.
Step 1 of 4
A = ∫20(x² + 1) dx
= [x³3 + x]20
= (83 + 2) − 0
= 143 ≈ 4.67
- Drag a plot to read coordinates, roots and extrema.
- A Matrix & Linear Solver handles Ax = b, inverses and eigenvalues.
- Insert LaTeX with a live preview, or write a formula by hand on the math canvas.
Fig. 04 Voice
Ask out loud. Hear it answered.
Speech recognition is set to run on the device, and replies are read aloud by local text-to-speech with six downloadable Kokoro voices. Audio recordings can be transcribed too.
Hands-free conversation mode listens, answers aloud and lets you interrupt naturally.Pro
Local voices
Fig. 05 The shelf
A shelf of 88 models, one already inside.
A small Qwen3 0.6B starter model ships in the app, so the first chat needs no download. From there the catalog runs from 0.14 GB up to Mac-class giants, plus Apple's Foundation Models on compatible iOS 26 devices.
Own AI picks a starting model that fits your device. 25 of the 88 take images; 22 are under 1 GB.
Spine height follows the number of models per family. The largest entries are meant for iPad Pro or Mac-class memory, not a phone.
The checklist is the app's own; the two animations are its Lottie files.
Bring one home from the HubPro
Search Hugging Face for community MLX models. Each is checked against your device before anything downloads: architecture, files, chat template, quantization and memory. Afterwards, a load-and-reply self-test proves it actually runs.
- Import a model folder straight from Files.
- Compare Models runs one prompt on two models and shows the replies and speed side by side.
- Automatic routing with a Faster, Balanced or Best Quality preference.
Fig. 06–09 Under the page
The parts you never see.
Running a language model on a phone is mostly a question of memory. Four things the engine does about it, each sketched in motion.
07 Draft models
A small model guesses ahead.
For a few larger models, a much smaller one of the same family proposes a handful of tokens and the larger model checks them in one pass. The output is identical to running the main model alone.
08 Memory-aware prefill
Long prompts, in chunks that fit.
Reading a long prompt is when a phone is most likely to end the app. So it is read in chunks sized from the memory the app has left, each using at most half the headroom above the device's memory floor.
09 Session cache
Chats keep their place.
Leave a chat and the model's working memory of it is written to disk. Reopen it with the same model and the full context comes back, instead of a clipped transcript being read again.
Settings › Performance keeps per-model medians measured on your own device: time to first token, generation speed, prompt-processing speed and peak memory. Prompts and replies are never recorded.
Fig. 10 Study
Questions that come back when they are due.
A document becomes a study set: flashcards and self-assessed quizzes, each with its source quote. A remembered question returns after 1, 3, 7, 14 and then 30 days. A missed one stays due.Pro
- A question whose quote can't be found in its passage is rejected.
- Check Answers with AI grades a typed answer: Looks Correct, Partly Right or Not Quite. You still choose.
- Study sets are encrypted on the device, and reminders are local notifications.
Try it: grade the card and watch when it returns.
Fig. 11 Yours
Give it a temperament, and a tint.
Seven built-in personalities, each with its own prompt and temperature. Memory is off by default; switched on, it stores only your name or details you explicitly ask it to remember, on this device.
Temperature
Lower is steadier, higher is looser. Pro adds a Prompt Library for up to 20 saved personas.
Pick an accent. This page re-inks itself.
The Modern look — paper and ink, with a colour of your choice — adds seven accents and three paper tones, each with a matching Home Screen icon.
Paper
Fig. 12 Beyond the app
It reaches past the app.
“Ask Own AI”
Siri and Shortcuts
Say the phrase, or build with three Shortcuts actions.
- Ask Own AI
- Get Answer from Own AI
- Find AI Models
Apple Watch companion
Speak or tap a quick prompt. The iPhone does the thinking, and a short answer comes back to your wrist. watchOS 11 or later.
Three widgets
Model Status, Formula of the Day and Quick Actions, on the Home Screen.
The share sheet
Share text or a link from another app, and ask about it.
ENAsk anything
DEFrag irgendetwas
ESPregunta lo que quieras
FRDemandez n’importe quoi
Four languages
The interface is localized in English, German, Spanish and French.
Plates Eleven screens, straight from the app
Every screen here is the real one.
Colophon What it needs, what it costs
Download on the App Store.
- iPhone and iPad
- iOS 18 or later
- Local models
- A14 chip or neweriPhone 12, iPhone SE (3rd generation) or later
- Apple Foundation Models
- iOS 26and an Apple Intelligence-compatible device
- Apple Watch
- watchOS 11 or later
- Languages
- English, German, Spanish, French
- Source code
- GNU AGPL v3
Free
Chat on-device with a daily message allowance. Picking any catalog model is free.
Pro
Lifts the message and document limits. Offered monthly, yearly or as a lifetime purchase.
- Conversation Mode
- Study Sets
- Prompt Library
- Conversation Export
- Chat Folders
- Hugging Face Models
- Import Your Own Models
- Compare Models
- Image Input
What leaves the device, and when. Chatting with a local model sends nothing. Downloading a model fetches its files from Hugging Face, which can see your IP address and device headers, but no personal content. A web search sends the query you typed to DuckDuckGo by default, or to Google if you choose it. If you pick Apple Intelligence, prompts and document text may be processed by Apple, including Private Cloud Compute, and the app asks before that model is used.