apple-intelligence-foundation-models

To run on-device Foundation Models for text generation, you configure and manage a FoundationModels session within your iOS app. This approach processes prompts locally, enabling privacy-preserving AI generation without sending data to external servers.

Yes, Foundation Models support structured data generation alongside standard text output. You can enforce specific schemas and structured formats from your prompts, allowing your app to directly parse the on-device model's responses into usable data objects.

You enable tool calling with Foundation Models by integrating tool definitions directly into your Swift session. This allows the on-device model to autonomously invoke specified functions during generation, automating tasks and workflows locally on macOS.

Yes, you can check Foundation Model availability before starting an AI session. The framework provides session availability checks to verify that the device supports on-device AI, ensuring your iOS or macOS app handles environments lacking hardware capabilities gracefully.

Foundation Models support token-usage measurement and include built-in safety guardrails. You can monitor token consumption during prompting and generation, while the safety guardrails automatically filter inappropriate content during on-device AI processing.

Limitations of on-device Foundation Models include hardware availability constraints on older devices and token limits during generation. Because inference runs entirely locally, complex AI assistants may face performance boundaries compared to cloud-based server inference.