Apple's Foundation Models framework now drives Claude and Gemini through one Swift API, with free Private Cloud Compute for small developers
- Apple's Foundation Models framework extends past the on-device model to server models like Claude and Gemini through the same Swift API, and any provider can ship a Swift package conforming to the Language Model protocol.
- Developers enrolled in the App Store Small Business Program with fewer than 2 million first-time App Store downloads can run Apple Foundation Models on Private Cloud Compute with no cloud API cost.
- Dynamic Profiles let an app swap model, temperature, reasoning level, and toolset mid-session while one transcript carries across every switch, a building block for agents and skills.
- Image understanding accepts images alongside text prompts, with the Vision framework wired in for on-device OCR and barcode reading.
- Apple already lists third-party packages for Core AI models, MLX Swift, Claude, and Gemini (via Firebase), alongside an evaluations framework, a CLI, and open source components.
Hacker News 의견들
Why is this sitting near the top of HN, did something get updated? I remember it being discussed last year when it was announced, and there are rumors Apple is about to ship a new set of local models.
I've been using the on-device model in my terminal for correcting errors and suggesting commands, and it's surprisingly capable on an M1 with 8GB. That's about the smallest machine local AI is usable on right now.
Do the safety restrictions get in your way? They felt pretty aggressive for normal usage, and I wondered whether sticking to command line options keeps things tame enough that the model stays out of the way.
I'll just ask my LLM to build me a framework instead, thanks.
Can the local model actually code at all, like as the predictor in Zed?
Even the stronger Private Cloud Compute model struggled with the task I gave it, building a web-based side-scrolling platformer. I pointed Cursor at it using this pcc-server-example repo to test.
Claude has had App Attest support for months and I'm surprised Anthropic isn't promoting it more. Firebase was the practical option for calling LLMs straight from iOS apps without a fancy setup.
Curious whether you can use a non-API Claude or ChatGPT plan with this, that would actually be pretty useful.
I've been using Foundation Models in Noodle for a while and liked working with it. I haven't tried Private Cloud Compute yet because I think I need entitlements that only the Siri desktop app has for now.
So is the PCC thing free tokens? It really reads like free tokens.
Free samples of tokens, basically. You need an Organization developer account rather than an Individual one to get the entitlement, but after that it's easy, and I've been shipping it since WWDC. PCC usage is allotted to the end user's Apple ID, not my dev account, so I can ship a lightweight app without worrying about my own server costs.
Their Core AI work is flying under the radar. There's a repo at apple/coreai-models where they're adapting a ton of open weight models to run natively in Swift.