foundation-models

What it does

Provides a comprehensive guide to Apple's Foundation Models framework for on-device AI inference, structured output generation, tool calling, and streaming responses in iOS 26. It covers creating sessions, configuring parameters like maxTokens, temperature, topP, penalties, and examples of basic text responses. It also details the @Generable macro for structured outputs, complex nested data, and adding descriptions. Additionally, it describes Guided Generation with system instructions and constrained outputs, Tool Calling with defining tools and integrating them into a session, Streaming Responses with basic and structured output, Multi-Turn Conversations for maintaining context and explicit message history, Error Handling for common errors and timeouts, and SwiftUI integration including a complete chat interface and UI-friendly structured output.

How it works

• Import FoundationModels and create a LanguageModelSession (with optional Configuration(maxTokens, temperature, topP, frequencyPenalty, presencePenalty))
• Use session.respond(to: ...) to get text responses; access response.content for plain text
• Check availability via LanguageModelSession.isAvailable to determine Foundation Models support
• Use the @Generable macro to define structured outputs (e.g., MovieRecommendation, RecipeAnalysis, Recipe, TaskExtraction) and obtain typed results from session.respond
• Use systemInstructions to steer output in Guided Generation, or provide constrained outputs via appropriate Generable types
• Define Tools with @Tool and @ToolFunction, then invoke session.respond with tools to have the model call tool functions and incorporate results
• Stream responses using session.streamResponse(to: ...), handling partial content and optional structured updates via PartiallyGenerated<...>
• Manage multi-turn conversations by reusing the LanguageModelSession to preserve context or by supplying explicit message history
• Handle errors with specific catches (LanguageModelError.unavailable, .contentFiltered, .contextLengthExceeded, .cancelled) and implement timeouts using withTimeout pattern
• Integrate with SwiftUI for chat interfaces and render streaming or structured outputs in the UI

When to use it

• When building on-device AI capabilities on Apple devices without cloud dependency
• When you need privacy-first, offline-capable LLM interactions with low latency
• When you want type-safe, structured outputs for AI responses using @Generable
• When integrating tool calls (weather, restaurant data, etc.) into AI responses

What it can touch

• The framework is accessed via Swift imports (e.g., FoundationModels)
• Uses session creation and configuration objects, and optional tools passed to session.respond
• Demonstrates usage of @Generable-annotated structs/enums to constrain output

Caveats

• Requires iOS 26+, macOS Tahoe+, iPadOS 26+, visionOS 26+ and compatible hardware (A17 Pro+ or M-series) with Apple Intelligence support
• Inference runs locally with no data leaving the device; no API fees or cloud dependency
• Example code shows specific parameter defaults (maxTokens, temperature, topP, penalties) and may require adaptation for real apps
• Availability may vary by device capability; check LanguageModelSession.isAvailable before usage