Apple rebuilds Apple Intelligence on a Gemini foundation at WWDC 2026

Apple announced at WWDC 2026 that its third-generation Apple Foundation Models are co-developed with Google and distilled from a custom Gemini model, under a multi-year partnership Bloomberg’s Mark Gurman reports at roughly $1 billion per year for a 1.2-trillion-parameter base. The consumer-facing reveal followed a January 12, 2026 commercial deal that went underreported at the time. Apple says it evaluated OpenAI and Anthropic before selecting Google, framing the choice as a model-capability decision rather than a strategic alignment. The architecture preserves Apple’s runtime and privacy story while explicitly conceding that the in-house model program was not producing the capability gap Apple needed.

What shipped

The AFM 3 family ships as five models. AFM Core and AFM Core Advanced target on-device inference, with the Advanced variant offering a sparse, natively multimodal architecture. AFM Cloud and AFM Cloud Image run inside Private Cloud Compute for taxing requests and image generation. AFM Cloud Pro is the agentic and reasoning tier. Each model is trained on Apple Silicon with Apple’s data and then polished through distillation against the loaned Gemini model. AppleInsider’s reading, after access to Apple’s technical disclosure, is that the runtime artifacts are Apple code on Apple hardware. Google’s framing is consistent with that: the partnership provides a foundation, training was done on Google TPUs, and Google receives no Apple user data.

Apple’s own benchmarks claim AFM 3 Core was preferred over the 2025 baseline in 45.6 percent of human-rated general text comparisons, up from a 23.3 percent rate for the prior generation. The image, speech, and visual-question-answering capabilities are new. Siri’s personal-context and multi-step app-action features that Apple announced in 2024 and then repeatedly postponed are now being framed as the first features that require AFM 3’s capability ceiling.

A second-order announcement closes the developer loop. Apple’s Foundation Models framework now exposes a public LanguageModel protocol that third-party providers can implement, starting in iOS 27, macOS 27, iPadOS 27, visionOS 27, and watchOS 27. Google’s Gemini models are the first non-Apple option through that framework, and Gemini integration in Xcode 27 ships alongside it. The framework formalizes what was previously a partner-specific integration into a model-provider interface that Anthropic, OpenAI, and others can plug into on the same terms.

Where this lands in the market

Apple’s on-device model strategy from 2024 and 2025 was the only credible in-house frontier-model program at hardware companies its size, and the WWDC announcement is the clearest concession yet that capability is not the dimension Apple intends to compete on at the model layer. The runtime, privacy posture, and developer surface are the things Apple keeps. The model itself becomes interchangeable.

The OpenAI line, which began with Apple’s 2024 ChatGPT integration, is demoted by this announcement. ChatGPT remains a deferral target for Siri in 2026, but the Apple Intelligence base no longer sits on top of OpenAI’s capabilities, and the developer-facing model surface now treats Google, OpenAI, and Anthropic symmetrically. Anthropic, which has been Apple’s quietest model partnership to date, gets a credible on-platform path through the Foundation Models framework without needing the headline deal Google just secured.

For Google, the partnership is a routing arrangement as much as a model licensing arrangement. A custom 1.2-trillion-parameter Gemini shipping through hundreds of millions of Apple devices is a distribution channel Google has not had at this scale since Android’s earliest Gemini surfaces. The non-exclusive structure and a December 2025 antitrust ruling against exclusive default agreements between the two companies bound how far either side can lean into the relationship operationally.

What’s worth watching

1. Anthropic and OpenAI on the Foundation Models framework. Apple’s public LanguageModel protocol is the integration surface. The pace at which Anthropic and OpenAI ship adapters, and the consumer features Apple ties to each, will determine whether the framework is a real model marketplace or a Gemini-default surface with two satellite providers.

2. Siri AI feature rollout cadence. Apple has committed AFM 3 to the long-delayed Siri personal-context and screen-awareness features, with developer betas starting at WWDC and public betas in July. The first external benchmark of the on-device personalization claims is the October to December consumer-OS launch window. That is the test of whether the Gemini distillation closed the capability gap Apple’s in-house program could not.

3. PCC trust posture under a non-Apple base model. Private Cloud Compute was built around the verifiable claim that Apple controls the server-side model and runtime. AFM Cloud and AFM Cloud Pro distilled from Google’s model preserve that runtime story but add a training-time relationship Apple did not previously disclose. Researchers will probe how Apple documents the supply-chain trust boundary, especially the step at which Google’s contribution stops being visible to Apple’s PCC verifier.

Stackmaven’s follow-up on the framework adoption and on Siri AI’s first external test lands on or around September 7, 2026.

Sources cited

1. Google: Joint statement from Google and Apple blog.google
2. AppleInsider: Apple's new foundation models don't contain a drop of Gemini appleinsider.com
3. MacRumors: Apple Reveals New AI Architecture Built Around Google Gemini Models www.macrumors.com
4. Business Standard: WWDC 2026 Apple unveils Siri AI, Gemini-powered Apple Intelligence www.business-standard.com