AIHeroes 2026: Apple Foundation Models Workshop

Hands-on workshop building a single-screen iOS 26 chat app powered entirely by Apple's on-device Foundation Models framework — no network, no API keys, no per-token cost. Six checkpoints: availability, sessions, structured output, streaming, tool calling, and a typed transcript inspector. Presented at AI Heroes 2026, Turin. Repo: github.com/TParizek/AIHeroes-Apple-Foundation-Models-workshop

Transcript

1. Apple Foundation Models workshop
2. Today • Foundation Models introduction • Live coding the app • Takeaways • Q&A
3. 01 ON-DEVICE AI
4. Right on the phone • No network • No API key • No per-token cost • No data leaving the phone
5. LLM is a function • Text in • Text out • Surprisingly good at structure
6. 02 APPLE FOundation MODELS
7. Apple's on-device LLM • On device LLM with simple API • Free, offline, private • Minimum os: iOS 26, iPadOS 26, macOS 26 • Minimum processor: A17 Pro or any M series chip
8. THE interface • SystemLanguageModel • LanguageModelSession • @Generable • Tool
9. 03 Let’s build
10. Today's build • Availability • Sessions and instructions • Structured output and streaming • Tools and multi-turn
11. The repository • https://github.com/TParizek/AIHeroes-Apple-Foundation-Models-workshop • 7 branches: main → step-6-final • git diff is your answer key
12. Step 1 • ChatViewModel.swift ◦ Check Foundation Models availability & start session in init ◦ Update send function with session conversation • Test the app: expected result is a functional LLM
13. Step 2 • AssistantReply.swift ◦ Add @Generable and @Guide annotations ◦ Add “var sessions: [SessionCard]” to AssistantReply • ChatViewModel.swift ◦ Update session.respond to typed response • ChatView.swift ◦ Update func bubble -> case .assistant to render cards using SessionCardBubble • Test the app: expected result are visible session cards in responses
14. Step 3 • ChatViewModel.swift ◦ Add var isStreaming: Bool to AssistantMessage ◦ Update session.respond to session.streamResponse ◦ Handle partial updates of response • ChatView.swift ◦ Show TypingIndicator when isStreaming is true, but response is empty • Test the app: expected result is seeing response stream as it is generated
15. Step 4 • SearchSessionTool.swift ◦ Uncomment tool implementation ◦ Create SearchSpeakerTool.swift inspired by SearchSessionTool.swift • ChatViewModel.swift ◦ Update session creation with tools • Test the app: expected result is seeing response enriched with tool provided data
16. 04 Takeaways
17. What to remember • It's just a Swift API • Private by default • Free at runtime • Streaming, tools, multi-turn — all included
18. The hard part is done. Now build the easy part.
19. Questions?