The guide · Updated

How offline AI works on iPhone.

Offline AI means the model runs on your iPhone instead of a server. The app either ships with the model or downloads it once; after that, your iPhone's own chip does the work and the result is saved on the device. No upload, no round trip, and it keeps working in Airplane Mode.

Cloud AI compared with on-device AI

Cloud AI

  1. Your voice or prompt
  2. Uploaded over the internet
  3. A company's server runs the model
  4. Answer sent back

On-device AI

  1. Your voice or prompt
  2. Your iPhone runs the model
  3. Answer stays on the phone

Where does the model come from?

Either inside the app (ScribeKit's speech model ships with it) or as a one-time download you choose (Nashi's chat models, ScribeKit's optional Whisper models). Once it's on the phone, nothing more needs to be fetched to use it.

Why can an iPhone run AI at all?

Recent iPhones have chips designed for machine-learning work. Smaller, optimized models fit in the phone's memory and run fast enough for transcription and chat.

What can't on-device AI do?

Phone-sized models are smaller than the largest cloud models, so answers can be shorter and less knowledgeable. That's the trade for privacy, and for working anywhere.

How do I check an app is really offline?

Turn on Airplane Mode, turn off Wi-Fi, and use the app. Then look at App Privacy Report in Settings to see which domains it contacted.

The 5-minute Airplane Mode test, step by step

How our apps do it

Glossary

On-device
Processing that happens on your phone, not a server.
Model
The trained file that turns audio into text or a prompt into an answer.
Parakeet
NVIDIA's speech-recognition model, built into ScribeKit.
Whisper
OpenAI's open speech model, optional in ScribeKit Pro.
llama.cpp
The open-source engine Nashi uses to run chat models.