Apple's Foundation Models framework now extends to Claude and Gemini

Apple's Foundation Models framework now extends to Claude and Gemini

Apple's developer page for the Foundation Models framework describes it as a native Swift API for integrating any large language model into an app. That includes direct access to the Apple Foundation Models, both on-device and in Private Cloud Compute. Apple lists guided generation, tool calling, image understanding and Dynamic Profiles as built-in capabilities, and calls the framework "the best way to use any large language model in your app".

The headline change is reach beyond Apple's own models. The page says on-device models suit many tasks, but sometimes a larger model is needed for a more complex workflow. The framework now extends to server models, "like Claude, Gemini, and more", through the same Swift API, with full support for features such as tool calling and guided generation. The mechanism is the Language Model protocol: any model provider can create a Swift package that conforms to it, which lets a developer pick the right model for the app.

For on-device use, the framework gives direct access to the model at the core of Apple Intelligence, so apps can work privately and without an internet connection. Multimodal prompts let an app pass images alongside text for image understanding. The Vision framework is integrated directly, giving the model tools such as OCR for text extraction and barcode readers, all on-device.

There is a free tier for newcomers. Developers enrolled in the App Store Small Business Program who have fewer than two million first-time App Store downloads can use the Apple Foundation Models running on Private Cloud Compute with no cloud API cost. The sentence carries an asterisk whose footnote is not in the text. Apple says this delivers frontier-level intelligence with privacy protections, and that users' data is not stored or accessible to Apple or anyone else.

Dynamic Profiles are new declarative APIs for updating a session's model, tools and instructions on the fly. Instead of creating a session with a fixed configuration, a developer defines profiles, each with its own model, temperature, reasoning level and toolset, and swaps between them within one continuous session. The session transcript is shared across every profile switch, which Apple says gives the app more contextual intelligence with less prompting. Apple presents them as building blocks for AI agents, skills or other high-level abstractions.

The Evaluations framework is meant for testing prompts and checking that intelligence-powered features work across dynamic conditions, going beyond what unit tests alone catch. Apple pitches it for moments when a developer is changing instructions, switching models or tuning a Dynamic Profile.

The new fm command line tool in macOS 27 is for scripting and automating model workflows and prototyping without writing app code. A Foundation Models framework utilities package adds experimental abstractions: custom skills, context management helpers and a Chat Completions client that connects to a hosted model of the developer's choice. A Foundation Models SDK for Python extends the framework into the Python ecosystem, with integration into popular tooling and evaluation packages.

Finally, the page points to Acceptable Use Requirements that outline prohibited uses of Apple Foundation Models. Citing Article 53(1)(b) of the AI Act, Apple says it makes information and documentation available to providers of AI systems who intend to integrate AFM 3 Cloud into their systems.

Key facts

  • Apple says the Foundation Models framework, a native Swift API, now extends beyond its own models to server models such as Claude and Gemini, with tool calling and guided generation supported.
  • Any model provider can ship a Swift package conforming to the Language Model protocol, which is how developers pick the model for their app.
  • Developers in the App Store Small Business Program with fewer than two million first-time App Store downloads can use Apple Foundation Models on Private Cloud Compute with no cloud API cost (the asterisked footnote is not shown).
  • Dynamic Profiles let one session swap between profiles with their own model, temperature, reasoning level and toolset, while the transcript stays shared.
  • Tooling around the API includes an Evaluations framework, the fm command line tool in macOS 27, an experimental utilities package and a Python SDK.

Why it matters

Apple is positioning one Swift API as the way to work with any large language model in an app, not only its own. The page says the same calls, including tool calling and guided generation, now cover server models like Claude and Gemini as well as the on-device model behind Apple Intelligence. The Language Model protocol is the hook: any provider can publish a Swift package that conforms to it. Dynamic Profiles, which swap model, tools and instructions inside a single session, are Apple's building block for agents and skills.

Who it affects

Swift developers building apps that use language models are the main audience. Developers in the App Store Small Business Program with fewer than two million first-time App Store downloads are singled out for free access to Apple Foundation Models on Private Cloud Compute. Python developers get an SDK, and developers who script or prototype get the fm command line tool in macOS 27. Providers of AI systems who want to integrate AFM 3 Cloud are addressed in the AI Act paragraph.

How to use it

Build against the Foundation Models framework in Swift and choose the model through the Language Model protocol: Apple's on-device model, Apple's Private Cloud Compute models, or a server model supplied through a provider's Swift package. Pass images alongside text for image understanding; the Vision framework adds OCR and barcode reading on-device. Define profiles with Dynamic Profiles to change model, temperature, reasoning level and toolset mid-session. Use the Evaluations framework to test prompts before shipping, the fm tool in macOS 27 for scripting without app code, the experimental utilities package for custom skills, context management helpers and a Chat Completions client for a hosted model, and the Python SDK for Python tooling.

How solid is it

This is Apple's own developer page, so the features are described by the maker and read as product copy. Claims such as "best way", "frontier-level intelligence" and the privacy guarantee for Private Cloud Compute are Apple's statements, with no benchmarks or independent checks on the page. No release date or availability timeline is given on the page, and it does not say which Claude or Gemini versions are supported, nor whether Anthropic or Google have agreed to or shipped the Swift packages.

Risks and caveats

The free Private Cloud Compute offer has an asterisk whose footnote terms are not shown, and nothing says what happens above the two-million-download threshold. The page does not say whether server-model access is free, who bills it, or how third-party data handling works, so Apple's privacy statement should be read as applying to its own Private Cloud Compute. The utilities package is described as experimental. Apple's Acceptable Use Requirements set prohibited uses of Apple Foundation Models. No OS versions are named other than macOS 27.

“Any model provider can create a Swift package conforming to the Language Model protocol, a fundamental component of the framework that allows you to pick the right language model for your app.”

— Apple, Foundation Models framework developer page