> For those just getting started, developers who are enrolled in the App Store Small Business Program and have fewer than two million first-time App Store downloads can access the Apple Foundation Models running on Private Cloud Compute with no cloud API cost.*
> Where Apple Intelligence is available, eligible developers can use PCC in their apps distributed on the App Store, and test PCC features via TestFlight or ad hoc distribution. Installs during testing are not counted as first-time app downloads.
Is this free tokens? This sounds like free tokens?
Pretty much. You need an Organization developer account, not an Individual account to get the PCC entitlement, but after that it’s pretty easy. I’ve been using it in TestFlight and now a shipping app since WWDC and it’s pretty solid. A lot more performant than I had expected. Context windows could be a bit larger (though it’s much better than the on-device model). PCC usage is allotted to the end users Apple ID or iCloud account , not the app or my development account, which makes it possible to ship a lightweight app and not worry about my own server costs. Some of the background gotchas and PCC rate limiting behaviors are obnoxious but I’ll take what I can get. Some of it is just standard iOS behavior.
Yep. Though "free tokens" with unspecified daily limits is doing a lot of heavy lifting here. Still, for a small indie app that only needs occasional inference, this could be a pretty big deal
Not exactly. They’re not your tokens, they’re the user’s tokens. If those are different people.
Your app uses tokens when it goes to private cloud compute tied to the user’s iCloud account. I don’t think anyone knows what the limit is, we only know that higher iCloud plans bring higher limits.
It’s all very fuzzy right now as this is brand new and I think they’re just trying to incentivize development rather than risk strangling it out of the gate.
I actually think their CoreAI stuff is kind of flying under the radar. But they've got a repo at https://github.com/apple/coreai-models where they're adapting a ton of open weight models to their format to run natively with Swift.
Interesting. Have you had issues with the model safety restrictions at all? They seem rather aggressive for normal usage, I wonder if limiting to command line options keeps things safe enough they stay out of the way.
I have my firewall events get reviewed by the local model on my mac studio. I had to stop sending the full event because the model would refuse some requests.
I just realized that claude has also support App Attest for months [1]. Not sure if that is being used widely. Firebase was the most practicle option to call llms directly from ios apps without having fancy setups. I am surprised Anthropic not promoting this more.
I have been using Foundation Models in Noodle for a while and have liked working with it. One part I have not explored yet is Apple Private Cloud Compute but it is on my list. I think I need special entitlements that are only available to Siri desktop app for now - need to check.
Even the stronger Private Cloud Compute model struggled with the coding task I gave it (build a web-based side-scrolling platformer). I used this and pointed Cursor at it to test: https://github.com/littlebobert/pcc-server-example
We've reached the point where Apple is offering a unified abstraction over competing AI providers. Somewhere, a developer is already figuring out which provider-specific feature breaks the abstraction first.
https://developer.apple.com/private-cloud-compute/
> Where Apple Intelligence is available, eligible developers can use PCC in their apps distributed on the App Store, and test PCC features via TestFlight or ad hoc distribution. Installs during testing are not counted as first-time app downloads.
Is this free tokens? This sounds like free tokens?
Your app uses tokens when it goes to private cloud compute tied to the user’s iCloud account. I don’t think anyone knows what the limit is, we only know that higher iCloud plans bring higher limits.
It’s all very fuzzy right now as this is brand new and I think they’re just trying to incentivize development rather than risk strangling it out of the gate.
Surprisingly capable on an m1 8gb (in my opinion this is the smallest machine that local ai needs to run on)
[1] https://github.com/anthropics/ClaudeForFoundationModels#auth...
This is the repo if you are interested https://github.com/pdparchitect/noodle