Working with Apple Foundation Models

Apple has introduced the Foundation Models framework, a powerful new toolset that allows developers to integrate on-device large language models directly into their iOS and macOS applications. This framework is the engine behind Apple Intelligence, enabling features that are both smart and privacy-preserving.

What are Apple Foundation Models?

The Foundation Models framework provides access to Apple's own 3-billion-parameter language model. Unlike cloud-based solutions, this model runs entirely on your device. This approach offers significant benefits:

• Privacy First: Your data never leaves your device.
• Zero Cost: No recurring API fees for cloud processing.
• Offline Capable: Works without an internet connection.
• Low Latency: Immediate responses without network round-trips.

Getting Started

The framework is deeply integrated with Swift, making it incredibly easy to use. The core class you'll interact with is LanguageModelSession.

Checking Availability

Before using the model, you should check if it's available on the user's device. This requires the device to support Apple Intelligence.

import FoundationModels

if SystemLanguageModel.default.isAvailable {
    print("Foundation Models are ready to go!")
} else {
    print("This device does not support Foundation Models.")
}

Basic Text Generation

Generating text is straightforward. You create a session and simply ask for a response.

let session = try LanguageModelSession(model: .system)
let response = try await session.generate("Explain the concept of 'recursion' in programming.")
print(response)

Advanced Features: Guided Generation

One of the most powerful features is Guided Generation. Instead of parsing raw string output, you can ask the model to generate structured data that matches your Swift types.

import FoundationModels

@Generable
struct Recipe {
    let title: String
    let ingredients: [String]
    let steps: [String]
}

let session = try LanguageModelSession(model: .system)
let recipe: Recipe = try await session.generate("Create a simple pancake recipe.")

Using @Generable and @Guide macros, the framework ensures the model's output conforms perfectly to your struct, eliminating the need for complex prompt engineering and JSON parsing.

Implemented in Rider Weather

I have already started leveraging these capabilities in my own projects. In Rider Weather, I'm exploring ways to use on-device intelligence to provide even more personalized weather summaries and riding advice based on complex local conditions.

What's Next?

I have many more ideas for implementing Apple Foundation Models across my apps. The potential for "Tool Calling"—where the model can intelligently invoke functions within your app—opens up a new world of possibilities for natural language interfaces.

Stay tuned for more updates as I dive deeper into building with Apple Intelligence!