Skip to content
Riff

Written with Private Cloud Compute

Riff transcribes speech on your Mac. The writing, which is Polished cleanup, edit by voice, meeting notes, and answers to your questions, is done by Apple’s server model on Private Cloud Compute. Riff sends it the text a task needs, never audio.

What Apple says about Private Cloud Compute

Apple built Private Cloud Compute to run requests from Apple devices on Apple’s servers without keeping them. As of September 2026, Apple says the following.

  • In WWDC26 session 241, Apple says “No prompts are ever stored”, and that independent researchers can verify the claim.
  • Apple’s Private Cloud Compute security overview says that “PCC deletes the user’s data after fulfilling the request, and no user data is retained in any form after the response is returned”, and that user data “is never available to Apple”.
  • The same session says the model has “a 32,000 token context window”. That fits the transcript of an hour-long meeting in one request.
  • Each person has a daily allowance of Private Cloud Compute requests, which is higher with iCloud+, according to the same session.

Riff has no account and no server of its own between your Mac and Apple.

The daily allowance

Settings under Models shows today’s allowance, and Apple’s offer to raise it when it runs low. When the limit is close, Riff polishes dictation with the model on your Mac and leaves the allowance for meeting notes.

When Private Cloud Compute cannot answer

When your Mac is offline, the daily limit is reached, or Private Cloud Compute is not available on your Mac, Riff runs the same request on Apple’s on-device model, which needs Apple Intelligence turned on in System Settings. A note written this way says “Written on this Mac”. The on-device model reads less at a time, so Riff writes long meetings in parts, one for each of your notes.

Basic notes without either model

Without a writing model, Riff still writes notes, labeled Basic notes.

  • Under each of your paragraphs, the transcript sentences that match it best from around the time you typed it.
  • Action items, found by phrases such as “I’ll”, “can you”, “let’s”, and “follow up”.
  • A summary from the first sentence of each topic.

Dictation keeps working too. Polished cleanup falls back to Light, and edit by voice needs one of the two models.

Not used for training

Riff uses nothing you say, type, or record to train any model.