# Foundation Models

> On-device AI with Apple FoundationModels framework: text generation, streaming responses, structured output with @Generable. Use when adding on-device AI features, text generation, or ML-powered content creation. Triggers: FoundationModels, @Generable, LanguageModelSession, on-device AI, text generation.

- Skill: `abdullah4ai/foundation-models` (Agent Skill)
- Install (CLI): `npx skillmds@latest add abdullah4ai/foundation-models`
- Raw SKILL.md: https://api.skillmd.com/api/skills/abdullah4ai/foundation-models/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: abdullah4ai (https://skillmd.com/u/abdullah4ai)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/abdullah4ai/foundation-models

---

# Foundation Models

APPLE ON-DEVICE AI (FoundationModels — iOS 26+):
FRAMEWORK: import FoundationModels

AVAILABILITY CHECK (MANDATORY — model may not be available on all devices):
guard case .available = SystemLanguageModel.default.availability else {
    // Show "This feature requires Apple Intelligence" message
    return
}

BASIC TEXT GENERATION:
let session = LanguageModelSession()
let response = try await session.respond(to: "Summarize this text: \(userText)")
print(response.content)

STREAMING GENERATION:
let stream = session.streamResponse(to: prompt)
for try await partial in stream {
    displayText += partial.text
}

STRUCTURED OUTPUT with @Generable:
@Generable
struct RecipeSuggestion {
    @Guide(description: "Name of the dish") var name: String
    @Guide(description: "Estimated prep time in minutes") var prepTime: Int
    @Guide(description: "Main ingredients") var ingredients: [String]
}

let session = LanguageModelSession()
let recipe: RecipeSuggestion = try await session.respond(
    to: "Suggest a quick pasta dish",
    generating: RecipeSuggestion.self
)

SESSION INSTRUCTIONS (system prompt):
let session = LanguageModelSession(instructions: "You are a helpful cooking assistant. Keep responses concise.")

GUARDRAILS:
- Model output is filtered by Apple's safety system
- No internet required — fully on-device
- Context window is limited (~4K tokens typical) — keep prompts concise
- Use session.respond() for single turns, keep session for multi-turn conversations