AppleFoundationModels

On-device Foundation Models run a ~3 billion-parameter LLM directly on iOS, iPadOS, macOS, and visionOS, keeping all user data private and offline. Use LanguageModelSession with @Generable macros to ensure extracted data always conforms to a Swift struct, eliminating unreliable parsing.

Yes. Tool invocation extends the model to autonomously call custom app functions, allowing it to fetch real-time data or perform actions beyond built-in knowledge. This enables guided structured output generation in real-world app flows like summarization and entity extraction.

Constrained decoding using @Generable and @Guide macros ensures all model responses conform to specific Swift types. This eliminates string parsing errors and guarantees usable, structured outputs for classification, tagging, and data generation tasks.

Yes. The Content Tagging adapter produces structured labels for classification and tagging workflows. Foundation Models support these use cases across iOS, iPadOS, macOS, Mac Catalyst, and visionOS with full on-device processing and session-based interaction.

Foundation Models run entirely on-device with no external API calls required, ensuring offline functionality and user privacy. The framework handles natural language understanding and structured output generation using only device resources and LanguageModelSession.

Foundation Models excel at classification, tagging, summarization, and entity extraction with constrained decoding. Performance depends on device hardware; the ~3 billion-parameter model balances capability with on-device efficiency, best suited for guided generation rather than open-ended reasoning at scale.