foundation-models-ref

On-device Foundation Models let you run AI directly on iOS, macOS, iPadOS, and visionOS without server calls. Start by importing FoundationModels and creating a LanguageModelSession to access the 3B-parameter model with 4096-token context, full offline capability, and privacy guarantees built in.

The @Generable macro uses constrained decoding to guarantee type-safe Swift objects from the model. It enforces schema validation during generation, eliminating JSON parsing errors and ensuring the model output matches your expected data structure exactly.

Yes, the Tool protocol lets the model autonomously execute custom code during generation. You define tools that access WeatherKit, MapKit, Contacts, or your own APIs; the model decides when to invoke them to answer user queries without leaving the device.

Foundation Models excel at text summarization, classification, and entity detection but aren't optimized for world knowledge queries, complex multi-step reasoning, mathematical computation, or translation. Handle context limits with prewarming and error handling for exceededContextWindowSize and guardrailViolation.

Stream responses token-by-token using LanguageModelSession's streaming APIs to provide real-time feedback and manage context windows effectively. Prewarm sessions before critical operations and implement error handling for generation failures specific to on-device constraints.

Yes, Foundation Models distinguishes between prompts (user input) and instructions (system guidance). Use instructions to shape model behavior and context; prompts carry the actual user query. This separation enables consistent, controllable generations across use cases like tagging, extraction, and content generation.