Apple opened Foundation Models to cloud AI providers at WWDC 2026. Firebase AI Logic is now the Gemini entry point for iOS 27, in preview.
Apple unlocked its Foundation Models framework at WWDC 2026 (June 8–12), opening the LanguageModel
protocol to third-party cloud providers. Firebase AI Logic shipped the integration: Gemini cloud models that plug directly into the same Swift API your app already uses for on-device inference, with no extra backend required.
Key Takeaways #
- Apple introduced the public
LanguageModel
protocol atWWDC 2026, targeting iOS 27, macOS 27, iPadOS 27 and watchOS 27. - Firebase AI Logic now runs Gemini through the Foundation Models framework. SwiftUI views,
@Generable
structs and tool definitions stay unchanged (Firebase blog, June 2026). - Requires
Firebase SDK for Apple platforms v12.5.0+, currently on thewwdc26-preview
branch (Firebase docs, June 17, 2026). - Firebase App Check handles attestation automatically. Calls from tampered devices and scripts are rejected before reaching the model.
- This is a preview targeting iOS 27 / macOS 27 betas. Not production-ready today.
What Apple opened up #
Foundation Models in iOS 26 was on-device only. iOS 27 adds a public LanguageModel
protocol: a Swift interface any third-party provider can implement to expose their model through the same API surface as Apple’s built-in model. Your SwiftUI views, @Generable
structs and tool definitions compile against the protocol, not against a specific vendor. Swapping providers is a one-line change.
Google and Anthropic both shipped Swift packages implementing the protocol at WWDC. Firebase is Google’s production layer for it: App Check, project wiring and Firebase AI Logic’s existing SDK infrastructure, routed through the Foundation Models surface.
What Firebase adds #
Firebase AI Logic implements LanguageModel
for Gemini. Apple’s on-device model offers maximum privacy, no network round-trip and offline support. Gemini via Firebase brings large context windows, multimodal input and stronger reasoning at the cost of a network hop. The practical split: on-device for latency-sensitive local work, Gemini for anything that needs depth or vision. Firebase App Check sits in the chain with no extra configuration. Every call from an Apple device goes through attestation — the same mechanism Firebase uses for Auth and other services. Tampered binaries and emulated calls are rejected before the model sees the request.
Where the preview stands #
The Firebase docs (last updated June 17, 2026) list which Gemini capabilities surface through the LanguageModel
path versus the full Firebase AI Logic SDK. Streaming, function calling and multimodal inputs each carry caveats. Check the capabilities page before building.
This does not replace the standard Firebase AI Logic SDK path. Apps targeting iOS 26 or earlier, or requiring capabilities not yet in the protocol, still use Firebase AI Logic directly.
Getting started #
Add the Firebase SDK for Apple platforms v12.5.0+ via the wwdc26-preview
package branch. Add FirebaseAILogic
and FirebaseAppCheck
to your app target. The getting-started guide covers full setup including App Check configuration for the attestation provider.
iOS 27 beta is the requirement today. Once it ships, the protocol layer means on-device and cloud inference swap at runtime, model upgrades land independently, and app logic stays unchanged.