FoundationModels — On-Device LLM

Apple's framework for on-device generative AI. No cloud, no API keys, full privacy.

Critical Constraints

• ❌ DO NOT use response.output → ✅ Use response.content to access generated values
• ❌ DO NOT skip availability check → ✅ Always check SystemLanguageModel.default.availability first
• ❌ DO NOT exceed 4,096 tokens per session → ✅ Break large tasks into multiple sessions
• ❌ DO NOT send concurrent requests on same session → ✅ Check session.isResponding first
• ❌ DO NOT confuse with OpenAI/Anthropic APIs → ✅ This is Apple's native framework, different API surface

Availability Check (Required)

import FoundationModels

let model = SystemLanguageModel.default

switch model.availability {
case .available:
    // Show AI features
case .unavailable(.deviceNotEligible):
    // Device doesn't support Apple Intelligence
case .unavailable(.appleIntelligenceNotEnabled):
    // User needs to enable in Settings
case .unavailable(.modelNotReady):
    // Model downloading or not ready
case .unavailable(let other):
    // Other reason
}

Basic Session & Response

let session = LanguageModelSession()
let response = try await session.respond(to: "What's a good month to visit Paris?")
print(response.content)  // ← ALWAYS .content, never .output

Session with Instructions

let instructions = """
    You are a cooking assistant.
    Provide recipe suggestions based on ingredients.
    Keep suggestions brief and practical.
    """
let session = LanguageModelSession(instructions: instructions)
let response = try await session.respond(to: "I have chicken, rice, and broccoli")
print(response.content)

Guided Generation (@Generable)

Receive structured Swift data instead of raw strings.

@Generable(description: "Profile information about a cat")
struct CatProfile {
    var name: String

    @Guide(description: "The age of the cat", .range(0...20))
    var age: Int

    @Guide(description: "One sentence personality profile")
    var profile: String
}

let session = LanguageModelSession()
let response = try await session.respond(
    to: "Generate a cute rescue cat",
    generating: CatProfile.self
)
print(response.content.name)     // ← .content, not .output
print(response.content.age)
print(response.content.profile)

Collection with Count Constraint

@Generable
struct CookbookSuggestions {
    @Guide(description: "Cookbook Suggestions", .count(3))
    var suggestions: [String]
}

let response = try await session.respond(
    to: "What's a good name for a cooking app?",
    generating: CookbookSuggestions.self
)
print(response.content.suggestions)

Snapshot Streaming

Stream partially-generated structured output. @Generable produces a PartiallyGenerated type with optional properties.

@Generable
struct TripIdeas {
    @Guide(description: "Ideas for upcoming trips")
    var ideas: [String]
}

let session = LanguageModelSession()
let stream = session.streamResponse(
    to: "What are some exciting trip ideas?",
    generating: TripIdeas.self
)

for try await partial in stream {
    // partial.ideas is [String]? — fills in as tokens generate
    print(partial)
}

SwiftUI Integration with Streaming

struct StreamingView: View {
    @State private var partial: TripIdeas.PartiallyGenerated?

    var body: some View {
        VStack {
            if let ideas = partial?.ideas {
                ForEach(ideas, id: \.self) { Text($0) }
            }
        }
        .task { await streamIdeas() }
    }

    func streamIdeas() async {
        let session = LanguageModelSession()
        let stream = session.streamResponse(
            to: "Trip ideas for 2025",
            generating: TripIdeas.self
        )
        do {
            for try await snapshot in stream {
                partial = snapshot
            }
        } catch { print(error) }
    }
}

Tool Calling

struct RecipeSearchTool: Tool {
    struct Arguments: Codable {
        var searchTerm: String
        var numberOfResults: Int
    }

    func call(arguments: Arguments) async throws -> ToolOutput {
        let recipes = await searchRecipes(term: arguments.searchTerm, limit: arguments.numberOfResults)
        return .string(recipes.map { "- \($0.name): \($0.description)" }.joined(separator: "\n"))
    }
}

let session = LanguageModelSession(tools: [RecipeSearchTool()])
let response = try await session.respond(to: "Find me some pasta recipes")

// Error handling
do {
    let answer = try await session.respond("Find a recipe for tomato soup.")
} catch let error as LanguageModelSession.ToolCallError {
    print(error.tool.name)
    print(error.underlyingError)
}

Generation Options

let options = GenerationOptions(temperature: 2.0)  // Higher = more creative
let response = try await session.respond(to: prompt, options: options)

Session Transcript

let transcript = session.transcript  // View model actions during session

Context Limits

• 4,096 tokens per session (~12K-16K English characters)
• Instructions + prompts + outputs all count
• For large data: break into chunks across multiple sessions
• Error on overflow: LanguageModelSession.GenerationError.exceededContextWindowSize

Common Mistakes & Fixes

Mistake	Fix
response.output	response.content — always use content
Skipping availability check	Always check SystemLanguageModel.default.availability
Sending request while session is busy	Check session.isResponding first
Expecting cloud model quality	On-device model is smaller; keep prompts focused and simple
Trying to use on Simulator	Requires Apple Silicon device with Apple Intelligence enabled

References

• Generating content with Foundation Models
• Guided generation
• Tool calling
• WWDC 2025: Meet the Foundation Models framework