The most important Foundation Models change at WWDC26 is not merely a newer on-device model. Apple is expanding the framework from an API for one system model into a common model abstraction. The new LanguageModel protocol lets a LanguageModelSession work with Apple’s on-device model, Private Cloud Compute, Core AI, MLX, or another provider implementation.

Xcode 27 and iOS 27 are in beta, and these APIs should be treated as preview interfaces rather than stable deployment contracts.

A shared session is not shared capability

A common interface can reuse prompts, transcripts, tools, and generation configuration. It does not make backends equal. Models differ in modality, context, latency, cost, privacy path, authentication, and reasoning controls. An application should query availability and declared capabilities before selection. Text support does not imply vision, and vision support does not imply identical tool or context behavior.

Write selection as product policy. Prefer the on-device model for offline operation, low latency, and sensitive data when it meets the task. Consider Private Cloud Compute for workloads needing its larger remote model and reasoning. Integrate an organization-approved third party through a provider adapter. Record the selected model and configuration in the trace so that an output can be investigated later.

More than one model path

Apple presents a rebuilt on-device system model with image understanding. It also introduces PrivateCloudComputeLanguageModel; the WWDC session describes a 32K context window and selectable reasoning levels. Core AI and MLX provide paths for models shipped or downloaded by an application. These are deployment and governance alternatives, not quality-equivalent toggles.

The session also previews Swift packages from external providers. “Coming” must not be documented as installable on June 10. Verify the provider’s actual package and availability before shipping. Credentials belong in Keychain or a backend, never inside model instructions. Provider billing and service terms still apply even when the session API looks uniform.

Images, system tools, and profiles

The new model can receive image attachments in a prompt. Apple supports several image representations and arbitrary dimensions, while warning that larger images consume more tokens and latency. Normalize orientation and strip unnecessary metadata, but do not claim that every image needs one fixed crop. Test the types users actually supply.

New system tools include Vision-backed OCR and barcode reading, plus Spotlight-backed local retrieval. A native tool result is still data that needs validation. OCR can be wrong, and a barcode can contain an untrusted URL. Keep authorization outside the model and apply the same destination policies used for custom tools.

Dynamic Profiles can package tools, instructions, and model configuration for an agentic task. Treat a profile as versioned application configuration, not an object that user text may rewrite. A model change must never silently expand authority.

Evaluation becomes a first-class workflow

A common abstraction is especially useful when the same task suite can run against several configurations. Include ordinary inputs, unavailable devices, context overflow, failed image decoding, tool timeout, and refusals. Measure task quality, time to first token, total latency, token use, energy or remote cost, and the privacy route.

Do not compare only prose. Verify that structured results keep domain constraints, tool calls retain approval requirements, and session recovery behaves as expected. The common interface reduces wiring; tests must establish semantic portability.

A beta migration strategy

Place the new API behind a small adapter and feature flag. Keep an iOS 26 fallback where the product requires it. Use availability checks rather than platform-name assumptions. Persist your own versioned domain objects instead of a beta response representation that may change. Rebuild and replay image, tool, and session tests after every Xcode beta update.

Separate provider-specific metadata from the stable domain result. If a configuration exposes reasoning tokens or cache details, store them in an optional trace envelope. Do not force every provider to invent the same values. Unknown fields should remain inspectable without becoming mandatory application state.

WWDC26 makes Foundation Models look more like a model runtime layer than a convenience SDK for one model. That increases what Swift applications can compose across local, private-cloud, and third-party inference. It also moves model-selection responsibility into the application. A sound design does not promise that models are interchangeable; it uses capability checks, policy, evaluation, and permission boundaries to make each substitution testable.