foundation-models-on-device

On-device text generation in iOS uses Apple's FoundationModels framework to create a language model session directly on the device. You must check model availability before session creation to ensure offline, privacy-preserving AI features.

Guided generation uses the @Generable attribute in Swift to produce structured data types instead of raw strings. This ensures the on-device LLM returns parsed Swift types, enabling reliable integration of generated content into iOS applications.

Custom tool calling allows the on-device LLM to invoke your domain-specific Swift code during generation. This enables the model to execute app functions dynamically, extending Apple Intelligence capabilities beyond standard text generation.

Snapshot streaming provides real-time UI updates by streaming partially generated structured responses from the on-device LLM. This mechanism allows iOS applications to render text and structured data progressively as the model generates it.

Yes, the FoundationModels framework requires iOS 26 or later. Developers must implement availability checks before attempting to create a language model session to prevent runtime errors on unsupported devices.

On-device LLM session initialization fails when the device does not meet the iOS 26+ requirement or lacks Apple Intelligence support. Always perform availability checks before session creation to handle unsupported hardware gracefully.