Intelligence,
where you are.
A model that runs on your device, knows your context, and keeps both to itself.
In development. Nothing described here has shipped.1
What it is for
Six things a personal model should be good at.
Reasoning, locally
Work a problem through with a model running on the machine in front of you, with no request leaving it.
Your own context
Answers grounded in your mail, notes and files — read where they already are, not uploaded to be read.
Voice
Speech understood on the device, so being understood does not cost you a recording.
What is on screen
Ask about whatever you are looking at, without a screenshot going anywhere.
Writing
Draft and revise in your own voice, close enough to keep up with typing.
Actions
Ask for something to be done across your apps, and have it done rather than described.
In use
What you would actually ask it.
Reply to Sam, using the three pricing points from Tuesday's notes.
Reads Notes, Mail · stays on the device
What actually has to happen before Thursday, given what is already booked?
Reads Calendar · stays on the device
The photo from the roof in Lisbon. It was raining, and it was evening.
Reads Photos · stays on the device
What changed in this contract since the version they sent last month?
Reads Files · stays on the device
Why did the build get slower after Friday's commit?
Reads Projects · stays on the device
Speed
The fastest network is no network.
186,000 Miles per second. The speed of light, and the ceiling on how fast any answer can return from somewhere else.2
Where it runs
One family. Three places it lives.
On device
The default. The model lives on the machine you are holding and answers without a network.
Works offline. Nothing leaves. No cost per request.
On your desk
A larger model where there is room for one, for work a handheld cannot hold.
Longer context. Heavier reasoning. Still yours.
Asked first
If a request genuinely cannot be answered locally, you are told before it goes, and you can say no.
Explicit consent. Not retained. Declinable.
Privacy
What never leaves cannot be collected.
Processed where you are
Personal context is read on the device it already lives on. Being understood should not require uploading yourself first.
Not kept, not trained on
What you ask is not retained to improve a model. If that ever needs an exception, it gets asked for.
Told before it leaves
If something genuinely cannot be answered locally, you hear about it first, and you can decline.
Research
The open problems.
We have published nothing yet, so this is not a list of papers. It is what we are stuck on.
Capable models that fit in a pocket
Compressing a model until it runs on a handheld is easy. Doing it without hollowing out what made it worth running is the actual problem.
Designing the chip around the model
General-purpose parts force general-purpose software. What changes when the silicon is shaped to the thing it has to run.
Personal context without collection
Grounding answers in someone's own material while ensuring that material never becomes a dataset, including ours.
Usefulness, not benchmark scores
A model that tops a leaderboard and irritates the person using it has failed. The second measurement is the interesting one.
Developers
Build on it.
One surface across the hardware, the software and the models. A model on the device means features that work offline and cost nothing per request.