At WWDC 2026, Apple showed off a rebuilt version of Siri. The assistant runs on foundation models developed with Google. For complex queries, it taps Nvidia GPUs.
Apple used this year's Worldwide Developers Conference to unveil the next generation of Apple Intelligence and a ground-up rework of Siri. Marketed as "Siri AI," the new assistant can execute system-wide actions, read on-screen content, and pull personal context from messages, emails, and photos, according to Apple. A dedicated Siri app syncs conversations across all devices via iCloud.
Apple developed its "Apple Foundation Models" in close collaboration with Google, building on Gemini technology. During a Tech Talk after the keynote, Craig Federighi clarified just how far the partnership actually goes: "The amount of the Google Assistant we use is none." Apple doesn't use the Gemini app, the models Google ships to its own customers, or Google Search as a knowledge base. For world knowledge, Apple relies on its own "World Knowledge Service," which the company says it built over several years.
The third generation of AFM consists of five models: The four smaller ones - AFM Core, AFM Core Advanced, AFM Cloud, and AFM Cloud Image - were trained entirely for Apple Silicon and only "refined using outputs from Gemini frontier models," said Apple's AI lead Amar Subramanya. Only the top-tier model, AFM Cloud Pro, actually runs on Nvidia GPUs inside Apple's Private Cloud Compute infrastructure, which has been expanded into Google Cloud for that purpose. Subramanya told CNBC that AFM Cloud Pro is comparable in quality to Google's Gemini frontier models. No benchmarks have been released yet.
A "System Orchestrator" decides whether a query gets processed locally or in the cloud. Federighi called this router central to the system's privacy architecture.
The camera module gets a new Siri mode too. Users can point at a restaurant receipt to split the bill with Apple Cash or aim at a plate to get nutritional info. Visual Intelligence is also coming to visionOS.
Apple Intelligence will also be able to change compromised passwords automatically. The agent navigates through the relevant website, logs in, and saves the new password in the Passwords app. Apple hasn't said how reliably this works in practice.
Other app integrations include Safari with "Notify Me" and tab organization, a new Image Playground app, Spatial Reframing in Photos, AI video descriptions for HomeKit Secure Video clips, and summarized notifications in the Home app.
For developers, Apple is opening up the Foundation Models Framework, which now supports image inputs, custom skills, and server-side model execution. Xcode gets an expanded coding assistant for agentic workflows.
The most capable on-device model requires an iPhone 17 Pro or iPhone Air, an iPad with M4 and at least 12 GB of RAM, or a Mac with M3 and 12 GB of RAM. The standard iPhone 17 doesn't qualify with its 8 GB. Neither does the iPhone 16 Pro Max. Without the right hardware, many features are only available through Private Cloud Compute with higher latency.
Geographically, Siri AI is limited to English-speaking markets for now. In the EU, the feature won't launch on iOS or iPadOS at all. Apple blames the Digital Markets Act in a dedicated Newsroom post. Apple's reading of the regulation: the company would have to grant any third-party virtual assistant the same access Siri AI has, like reading and sending messages, making purchases, accessing files, and triggering actions across all apps. Security researchers have already shown that AI systems like these can be hijacked to steal passwords and photos or change account and file settings without consent.
As a compromise, Apple proposed a "Trusted System Agent" - a middleware layer that would securely give third-party assistants the same capabilities, paired with an 18-month phased rollout. The European Commission rejected that proposal, along with every other one Apple put forward. Apple hasn't given a timeline for bringing the feature to iPhone and iPad in the EU.
macOS 27 and visionOS 27 will get Siri AI in the EU, though. That's because the DMA's scope only designates iOS and iPadOS as core platform services of a gatekeeper. macOS and visionOS aren't designated, so they aren't subject to the interoperability rules that would force Apple to open up to rival assistants. The Apple Watch is collateral damage: watchOS 27 requires a paired iPhone with Siri AI active, which EU users won't have. EU developers can't test or integrate the new Siri AI features on iOS, iPadOS, or watchOS either.
Almost two years ago, Apple announced a personalized Siri with screen context and app actions at WWDC 2024, originally planned for iOS 18. The capabilities shown now - arriving very late and after major internal power struggles - largely cover what Apple demoed back in 2024. Tim Cook, who steps down as CEO in September and hands the role to John Ternus, opened the keynote one last time.
Apple's executives made a point of drawing a line between their strategy and the competition's. "Some appear to be racing forward, seemingly pursuing AI for the sake of AI, without clear regard for the people that it's ultimately meant to serve," Federighi said. Apple is positioning itself as the counterpoint to OpenAI, Anthropic, and Google: instead of pouring money into its own frontier models and massive data centers, the company is betting on a mix of licensed model expertise (Gemini technology), its own silicon optimization, and the privacy story around Private Cloud Compute.
Federighi told CNBC that unlike web-based assistants, Apple collects very little user data. Instead, it uses locally stored information like calendars and messages for personalization.
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。