Esaldgut AI Native Engineering Workspace Swift Ondevice AI Language Model Patterns

On-device AI with Foundation Models (iOS 26)

Apple's Foundation Models framework (WWDC25) exposes the on-device LLM behind Apple Intelligence to your app. You always gate on availability first, then run a LanguageModelSession, get type-safe output via guided generation (@Generable + @Guide), stream partial results, and optionally extend the model with the Tool protocol. This is a hardware-gated capability, not just a UI pattern — the framework leads with availability for a reason.

When to invoke

• You're adding on-device generation/understanding: summarization, content tagging, smart search, suggested replies, structured extraction — and want it private and offline.
• You need structured output (a Swift type, not a raw string) from the model.
• You must fail soft on devices without Apple Intelligence.

Announce on invoke: "Using swift-ondevice-ai-language-model-patterns to gate availability, run a LanguageModelSession, and generate @Generable output."

Do not reach for this for tasks needing world knowledge, long context, or guaranteed availability — the on-device model is small (4,096-token window) and absent on older hardware. For those, a server model behind your own API is the right call.

The canonical APIs (verified)

The rules (load-bearing)

1. Gate on SystemLanguageModel.default.availability BEFORE constructing a session

Constructing/using a session on an unsupported device fails. Check availability and branch to a fallback. The unavailable reasons are actionable (prompt the user to enable Apple Intelligence vs. hide the feature on ineligible hardware):

switch SystemLanguageModel.default.availability {
case .available:
    // proceed
case .unavailable(.appleIntelligenceNotEnabled):
    // deep-link to Settings, or show "turn on Apple Intelligence"
case .unavailable(.deviceNotEligible), .unavailable(.modelNotReady):
    // hide the feature / degrade gracefully
}

2. Structured output = a @Generable type with @Guided fields

Annotate the type with @Generable; describe properties with @Guide. The framework uses constrained sampling so the model can't produce malformed output. Keep descriptions short — they consume the context window.

@Generable
struct SearchSuggestions {
    @Guide(description: "Suggested search terms.", .count(4))
    var terms: [String]
}

3. Streaming yields a ResponseStream of partial snapshots — handle progressive reveal

streamResponse(...) returns a LanguageModelSession.ResponseStream<Content>; iterate it for Snapshots where the generated content is filled in progressively (a PartiallyGenerated mirror with optional fields). Your UI must render half-filled state, then settle. Call .collect() to await the final value instead.

4. The context window is 4,096 tokens — split large work into sessions

Instructions + all prompts + all outputs share one 4,096-token budget. Exceeding it throws exceededContextWindowSize(_:). For long inputs, chunk the work and run each chunk in a new LanguageModelSession, then combine. Tool definitions also consume the window.

5. Tools must be Sendable; the model decides when to call them

A Tool carries a name + description (the model uses them to decide invocation) and a call(arguments:) whose Arguments are themselves @Generable. Tools run concurrently, so the protocol requires Sendable.

6. Wrap it in a @MainActor @Observable capability provider

Surface availability and streamed state to the View layer through an @Observable object on the main actor. The same provider can gate other Apple Intelligence features (Writing Tools, Image Playground) behind one availability check.

Canonical example

import FoundationModels

@Generable
struct Recipe: Sendable {
    @Guide(description: "The recipe title.")            var title: String
    @Guide(description: "Ingredients with quantities.") var ingredients: [String]
    @Guide(description: "Ordered preparation steps.")   var steps: [String]
}

@MainActor @Observable
final class RecipeGenerator {
    enum State { case unsupported(SystemLanguageModel.Availability), idle, streaming(Recipe.PartiallyGenerated), done(Recipe) }
    private(set) var state: State = .idle

    func start() {
        if case .unavailable = SystemLanguageModel.default.availability {
            state = .unsupported(SystemLanguageModel.default.availability); return   // fail soft
        }
        state = .idle
    }

    func generate(prompt: String) async {
        guard case .available = SystemLanguageModel.default.availability else { return }
        let session = LanguageModelSession(
            instructions: "You are a chef. Produce a concise, safe recipe."
        )
        do {
            let stream = session.streamResponse(to: prompt, generating: Recipe.self)
            for try await snapshot in stream {            // progressive reveal
                state = .streaming(snapshot.content)
            }
            let final = try await stream.collect()
            state = .done(final.content)
        } catch {
            // includes exceededContextWindowSize — start a new session with a shorter prompt
        }
    }
}

Decision aid: when NOT to / trade-offs

• Not for world-knowledge or long context. 4,096 tokens and a ~3B on-device model — use a server model when you need breadth or large inputs.
• Availability is a moving target at runtime. Battery, Game Mode, or a still-downloading model can flip it to .modelNotReady; re-check, don't cache "available" forever.
• Don't ship a feature that hard-fails on A16 and older. Those devices report .deviceNotEligible — design the fallback (hide, or route to a server) as a first-class path.

Related skills

• global-skills/apple/swift-clean-architecture-module-scaffold/SKILL.md — the @MainActor @Observable capability provider and Sendable Tool isolation.
• global-skills/apple/swift-feature-scaffold-mvvm-clean-arch/SKILL.md — wiring the streamed state into a feature's State enum.
• global-skills/meta/skill-pattern-freshness-audit/SKILL.md — re-verify Foundation Models symbols after each WWDC (the model + API surface is new and evolving).

Sources

• Foundation Models · LanguageModelSession · SystemLanguageModel · Generable · Tool · Generating Swift data structures with guided generation
• WWDC25 286 Meet the Foundation Models framework · 259 Code-along: Bring on-device AI to your app

Last verified: 2026-06-03 against Apple Developer docs (live). Resolved draft open questions: @Generable/@Guide confirmed real (Generable protocol page, sample uses .count(4)/.range(1...10)); streamResponse(to:generating:includeSchemaInPrompt:options:) returns ResponseStream<Content>; Tool is protocol Tool<Arguments, Output> : Sendable; the on-device context window is 4,096 tokens (Apple docs), correcting the ~2000 estimate. Re-check after: WWDC26, or by 2026-12-01. Decay risk: medium (new framework; model size and API surface may shift). Found a drift? Run /skill-pattern-freshness-audit apple.

Source: esaldgut/ai-native-engineering-workspace — distributed by TomeVault.