Karta
An agent that thinks on your phone.
Speak, and the work happens on the device in your hand. The model is on the phone, so your voice, your request and the answer all stay there. It is named after the Sanskrit grammatical term for the doer of an action, which is the whole job.
- Inference
- On device
- Network calls
- None
- Model
- Apple on-device, ~3B
- Requires
- iOS 26+
- Available
- Not yet
Where the work happens.
Every assistant you have used sends your words to a datacentre and waits. Karta does the same work on the processor already in your pocket. The difference is not speed. It is that there is no copy of what you said anywhere else.
When that actually matters.
Not as a principle. As four ordinary situations where every other assistant stops being useful.
- none
Underground
On the tube, in a basement, in a car park. The places you most want to capture a thought are the places with no signal.
- off
In the air
A flight is several hours of the best thinking time you get, and the one place a cloud assistant is dead weight.
- flaky
On bad hotel wifi
Not offline, just slow enough that a round trip takes long enough to break your train of thought.
- full
When it is nobody's business
A note about your health, your salary, someone else's secret. Things you would type but would not upload, even with a perfect signal.
What is true today.
Karta is not finished and there is nothing to download. Rather than a waiting list, here is the actual state of it.
Working
- Apple's on-device foundation model answers without a network
- Voice in, text out, with the transcription also on device
- Nothing is transmitted, which is the part that was never going to be negotiable
Not working yet
- It is slower than a cloud assistant, and on older phones considerably slower
- It gets things wrong that a much larger model would get right
- No release date, no TestFlight, no sign-up form pretending otherwise