foundation-models-on-device

To use Apple's on-device LLM for private text generation, initialize a LanguageModelSession within FoundationModels to handle generation without sending user data to external servers. This provides offline text generation directly on the device.

Yes, you can generate structured Swift models from language model output using the @Generable property wrapper. It creates typed Swift models with predictable fields and PartiallyGenerated types for safer parsing and direct UI integration.

FoundationModels supports tool calling and streaming for SwiftUI updates by defining app-specific tools and streaming PartiallyGenerated snapshots. This allows progressive UI rendering as the on-device model generates output.

Limitations of Apple's FoundationModels include a 4,096 token context limit, single-session concurrency rules, and strict on-device privacy constraints. You must also handle model availability checks before attempting generation.

On-device LLM generation is available offline for iOS apps, providing privacy and offline capability. It processes guided generation, tool calling, and streaming locally on devices running iOS 26 and later without external network calls.