Start a project

← All posts

WWDC 2026: Foundation Models Puts On-Device LLMs in Every iOS App

· Dracode · ios · apple-intelligence · developer-tools
Computer screen showing program code and an app running in a modern office workspace

The headline obscured three new developer frameworks

At WWDC 2026, Apple opened with Siri AI — a ground-up rebuild of Siri with a chatbot personality, cross-app awareness, and an independent app. Apple’s stock fell nearly 5% in the hours after the keynote, shedding roughly $240 billion in market cap. Investors, apparently, expected more.

For developers building on iOS 27, that headline obscured three frameworks that change the baseline for what you can ship: Foundation Models, a materially expanded App Intents with the new View Annotations API, and a revised Image Playground. There is also a breaking deprecation — ImageCreator — that will prevent apps from compiling if you miss it before the public release.

Foundation Models: Swift-native on-device LLM access

The Foundation Models framework is the most significant developer API Apple has shipped since SwiftUI. It gives any iOS 27, macOS 27, or visionOS 27 app direct access to Apple’s on-device foundation models — no server round-trips, no third-party SDK required.

The design centers on a LanguageModel protocol. Apps can target Apple’s built-in model, or any provider that ships a conforming Swift package. Apple is defining the interface, not forcing a single model. That interoperability matters: you can use Apple’s model for on-device latency-sensitive tasks and swap in a more capable provider for heavier workloads, within the same session.

Key capabilities in the framework:

  • On-device and Private Cloud Compute: inference runs locally when the model fits in RAM; larger requests route to Private Cloud Compute with no data stored
  • Multimodal prompts: text and images as joint input
  • Dynamic Profiles: swap models, tools, or system instructions mid-session — the mechanism for multi-agent flows that hand off between specialized contexts
  • Evaluations framework: a built-in testing harness for validating inference behavior across prompts and session states

The Xcode 27 beta shipped on June 8. Public API documentation is still stabilizing, but the architectural direction is clear: on-device inference as a first-class primitive, alongside URLSession for networking and CoreData for persistence.

App Intents gets View Annotations in iOS 27

App Intents has been the most underused surface in Apple Intelligence. iOS 27 extends it in two ways that are worth understanding before the public release.

First, entity schemas now feed into Spotlight’s semantic index. Content your app surfaces as entities — a record, a product, a scheduled event — becomes part of the personal context Siri AI reasons over when answering user questions, even when your app is not open. This is how Siri AI achieves personal context without reading your app’s database directly.

Second, the new View Annotations API maps SwiftUI views to App Intent entities. Once you annotate views, Siri AI gains on-screen awareness of your app’s content. A user can say “open the third one” while looking at a list in your app and Siri resolves the reference — because your views declared which entities they represent.

The intent schema side also drops the requirement to define specific trigger phrases. Siri’s language model discovers what an intent does from the schema itself and routes to it through natural language. That is a significant improvement over the iOS 18 model, where undiscoverable exact-phrase triggers were the main friction point for App Intents adoption.

For developers who build mobile products from scratch, this is the architecture pattern to start with in iOS 27 projects: define your entity and intent schemas before you write any Foundation Models calls.

Image Playground is a migration you cannot defer

The ImageCreator class is deprecated and will not compile in apps targeting iOS 27, iPadOS 27, macOS 27, or visionOS 27. During the beta, Xcode will issue warnings. Once the public OS ships, the build fails.

The replacement is the imagePlaygroundSheet modifier in SwiftUI, or ImagePlaygroundViewController in UIKit and AppKit. The new pattern is sheet-based: users interact with a system-provided generation interface, and your app receives the result in a completion handler.

.imagePlaygroundSheet(
    isPresented: $showingPlayground,
    concepts: concepts,
    onCompletion: { url in
        // url is a local file URL for the generated image
    }
)

Concepts take several forms — a .text() description, a .extracted(from:title:) extraction from longer context, a source UIImage for visual inspiration, or a PKDrawing from PencilKit. Apple also added an @Environment(\.supportsImageGeneration) key for capability checks at runtime.

All generation runs on Private Cloud Compute. Images are not stored or shared. If you are using ImageCreator anywhere in your codebase, this migration needs to start now. The beta window before public release is roughly three months.

The hardware floor and the EU gap

Not all of WWDC 2026’s features ship to all users on the same date.

Siri AI’s conversational mode — the cross-app agent behavior, extended context, and chatbot personality — requires 12 GB of unified memory. That limits it to the iPhone Air, iPhone 17 Pro, and iPhone 17 Pro Max. Devices with less RAM get a degraded Siri experience. Foundation Models and Image Playground run on the broader Apple Intelligence-compatible set, which includes hardware back to iPhone 15 Pro.

There is also a geographic split. Apple confirmed Siri AI will not launch in the European Union at release. The EU Commission responded that the decision is Apple’s alone. DMA compliance is the implied reason. For products targeting EU users, Siri AI integration is unavailable at launch — but Foundation Models, App Intents, and Image Playground still work.

These are not edge cases. EU and older-device users are mainstream segments for most products. Designing for the lower tier as the baseline — rather than as a fallback — is the right starting point.

What this means for the products we ship

The Foundation Models + App Intents pairing is how Apple intends AI-native iOS apps to be built. Foundation Models handles inference; App Intents handles the surface area Siri AI and Spotlight can access. They compose.

Our default architecture for new iOS 27 projects starts with entity and intent schema design first. That work is not Apple Intelligence-specific — it improves app quality regardless of whether the user has a 12 GB device or is in the EU. Foundation Models calls layer on top for features that genuinely need in-process inference. That order of operations keeps the app functional for every user and progressively richer for those on capable hardware.

The ImageCreator deprecation is the immediate action item. If you have an existing app using it, the migration is mechanical but it needs to ship before the public OS release. If you want to review the scope, reach out.

Sources

  1. Apple Intelligence — Developer Documentation — Apple, June 2026
  2. Deprecation of the ImageCreator class — Apple Developer News, June 11, 2026
  3. Create high-quality images using Image Playground — WWDC26 session 375 — Apple, June 2026
  4. iOS 27 beta and Xcode 27 beta — Developer Releases — Apple, June 8, 2026
  5. Apple Shares Massive List of Over 250 Changes Across iOS 27 and More — MacRumors, June 10, 2026
  6. EU Says Decision Not to Launch Siri AI in Europe Is Apple’s Alone — MacRumors, June 9, 2026