Siri AI at WWDC 2026: Apple Bets on Gemini and Vision LLMs Instead of App Integrations

Apple rebuilt Siri from scratch and licensed a custom Gemini-derived model to power it — running on Apple’s own Private Cloud Compute infrastructure rather than routing queries through Google.

The three things developers should actually track:

• Vision LLMs replace app integrations. Instead of requiring every app to ship custom Apple Intelligence hooks, Siri AI uses vision models to extract information directly from the user’s screen. As Simon Willison notes, this “neatly sidesteps” the integration problem that plagued the 2024 announcements. Vision LLMs were far less capable when Apple first pitched Apple Intelligence in June 2024.
• Core AI library ships with PyTorch support. A new developer framework called Core AI lets you run your own models on Apple silicon. It integrates with Meta’s open-source PyTorch ecosystem through dedicated Core AI PyTorch Extensions (coreai-torch), which is the first time Apple has built a first-party bridge to PyTorch rather than pushing developers exclusively toward Core ML.
• Conversational Siri across the OS. Per MacRumors, the update spans system-wide integration with a redesigned interface on iOS 27 and macOS 27 Golden Gate, supporting conversational input that can pull from both the public internet and your Apple account data.

The screen-reading approach is the smartest architectural choice here. It turns every existing app into an AI-accessible surface without any developer work — a pragmatic move after the 2024 promises went largely undelivered. Willison is holding to a “believe it when I see it” policy given how badly Apple’s previous Apple Intelligence timeline slipped, and that skepticism is earned.

The coreai-torch library is the immediate action item for developers building on-device AI. It finally gives Apple hardware a direct path from PyTorch training to local inference without a Core ML conversion step.