Apple ships a Gemini-trained Siri rebuild and four new Apple Foundation Models at WWDC 2026
Craig Federighi introduces Siri AI as a "profoundly more intelligent, knowledgeable, and capable Siri," wired across iPhone, iPad, Mac, Apple Watch and Vision Pro, with a dedicated app for conversation history and a Settings panel that lets users swap ChatGPT, Claude, Google Gemini or Grok in as the default model behind Apple Intelligence prompts. Four new Apple Foundation Models (AFM Core, AFM Core Advanced, AFM Cloud, AFM Cloud Pro) split between Apple Silicon Neural Engines and Private Cloud Compute servers carry the load. AFM Cloud Pro runs on Google Cloud-hosted NVIDIA GPUs under Private Cloud Compute attestation. Apple confirms the models were co-developed with Google through distillation and training but contain no Gemini code or infrastructure in production. Tim Cook opens the keynote at Apple Park at 10:00 PT for what is treated as his last WWDC as CEO, with John Ternus formally taking the chief executive seat September 1.
Apple Newsroom / TechCrunch / AppleInsider / Bloomberg · 2026-06-08 · Source ↗ ·
Apple ships a Gemini-trained Siri rebuild and four new Apple Foundation Models at WWDC 2026
Craig Federighi introduces Siri AI as a "profoundly more intelligent, knowledgeable, and capable Siri," wired across iPhone, iPad, Mac, Apple Watch and Vision Pro, with a dedicated app for conversation history and a Settings panel that lets users swap ChatGPT, Claude, Google Gemini or Grok in as the default model behind Apple Intelligence prompts. Four new Apple Foundation Models (AFM Core, AFM Core Advanced, AFM Cloud, AFM Cloud Pro) split between Apple Silicon Neural Engines and Private Cloud Compute servers carry the load. AFM Cloud Pro runs on Google Cloud-hosted NVIDIA GPUs under Private Cloud Compute attestation. Apple confirms the models were co-developed with Google through distillation and training but contain no Gemini code or infrastructure in production. Tim Cook opens the keynote at Apple Park at 10:00 PT for what is treated as his last WWDC as CEO, with John Ternus formally taking the chief executive seat September 1.
Apple Newsroom / TechCrunch / AppleInsider / Bloomberg · 2026-06-08 · Source ↗ ·
Tim Cook walked out at Apple Park at 10:00 PT Monday for what insiders treated as his last WWDC keynote as CEO. The succession story sat in the room from the first sentence; AppleInsider's pre-keynote running order, which surfaced over the weekend, had Cook opening, Federighi carrying Apple Intelligence, Mike Rockwell introducing the rebuilt Siri, Jeff Norris taking visionOS, David Clark on watchOS and Dr. Sumbul Desai closing Health. The September 1 handover to John Ternus, the senior vice president of Hardware Engineering Bloomberg's Mark Gurman has been reporting since May 25, was treated by the audience and the on-stage choreography as effectively decided. Cook stays on as board chair after the transition.
The technical centerpiece is Siri AI. Federighi framed the rebuild in his on-stage segment: Apple is "introducing Siri AI, a profoundly more intelligent, knowledgeable, and capable Siri" that can answer questions about on-screen content, search across messages, emails and photos using personal context, and reach to the web when the request needs current data. Siri now has a dedicated app with conversation history that syncs across devices via iCloud. The model behind the conversational layer is a custom Apple Foundation Model co-developed with Google through distillation from Gemini-family teachers. Apple emphasized at the keynote and in the developer-track sessions that no Gemini code, infrastructure or production weights sit inside the Apple stack. AppleInsider's Wesley Hilliard headlined the point Monday afternoon: "Apple's new foundation models don't contain a drop of Gemini."
Four models, two tiers
The Foundation Models lineup is now four named variants. AFM Core is the base on-device model that handles Writing Tools, summaries, Smart Reply, image-aware tasks and the lightweight requests that fit on Apple Neural Engine silicon. AFM Core Advanced is a natively multimodal model with a sparse architecture that runs more capable features on-device without leaving the user's phone or Mac. AFM Cloud is the Private Cloud Compute base model for taxing requests that exceed on-device budget, with a dedicated image generation and editing variant. AFM Cloud Pro is the heaviest model, deployed on Google Cloud servers and NVIDIA GPUs but still wrapped in the Private Cloud Compute attestation protocol so third parties can verify Apple isn't reading the traffic. The on-device tier is what Apple controls end-to-end; the cloud tier is what Apple pays Google roughly $1 billion a year to run.
"We use none of the models that Google deploys to their customers, nor do we use the infrastructure and means by which they deploy models to their customers." — Apple, WWDC 2026 Platforms State of the Union
Apple Intelligence Extensions and the user-picked default
The headline architectural decision is the Apple Intelligence Extensions framework. Users can set Claude, ChatGPT, Google Gemini or Grok as their preferred backend across Siri, Writing Tools, Image Playground, Visual Intelligence and the Apple Intelligence system features through Settings, with a dedicated App Store marketplace for adding third-party providers later. Gemini is the out-of-box default. Each external provider carries a distinct voice in spoken responses so users hear which model answered. The shift ends Apple's single-provider arrangement with OpenAI that shipped with iOS 18.2 and reframes the on-device Apple stack as a routing and privacy layer above a competitive model market.
Xcode 27 and Core AI
Xcode 27 ships with coding agents from Anthropic, OpenAI and Google built into the developer workflow. The agents handle test writing, multi-file refactors, interactive planning via a new /plan command, localization, String Catalog generation and translation across languages. Agent conversations live in the editor pane with full tab and split-window support. Xcode 27 is a dual-engine system: a local model running natively on Apple Silicon Neural Engines offers code and documentation suggestions tied to the active Swift and Apple SDK project structure in real time, while heavier code analysis and structural bug-finding routes to third-party frontier models. Apple's new Core AI framework gives developers direct API access to deploy full-scale large language models on Apple Silicon and leverage the Neural Engine for local processing. Apple Foundation Models framework — which exposed the 3B-parameter on-device Apple Intelligence base last year — gains image input alongside text and adds custom skills plus the ability to run jobs on Private Cloud Compute servers. Developers in the App Store Small Business Program with fewer than 2 million users get Private Cloud Compute access free.
Apple Susan Prescott, the VP of Worldwide Developer Relations, framed it for the developer audience: "With new intelligence frameworks and agentic coding in Xcode 27, developers have the tools they need to focus on what they do best." Federighi closed the privacy theme of his segment with a one-line restatement: "We believe privacy in AI is non-negotiable." The Private Cloud Compute attestation layer is the practical instantiation of that line — independent third parties can audit the server-side handling of user requests, including when AFM Cloud Pro fans out a query onto Google-hosted NVIDIA hardware.