Apple's new Siri runs on a custom Gemini and Google Cloud GPUs

Apple used its WWDC 2026 keynote to reset Siri, and the architecture Simon Willison pulled out of the announcements is the part worth pausing on. The new on-device experience is backed by a custom Gemini-derived model that Apple licensed from Google. Heavy tasks, what Willison calls "agentic tool-use and complex reasoning," run on Apple's Private Cloud Compute, which itself sits on Google Cloud with NVIDIA GPUs under the privacy and attestation wrappers Apple put in place two years ago.

For developers Apple shipped Core AI, a library for running models on Apple hardware that hooks into PyTorch through a Core AI extension. Existing PyTorch checkpoints can be converted to a format the on-device runtime understands. The Siri integration also leans on vision language models to read app screens directly, which Apple says removes the need for individual apps to expose new APIs to make their state visible to Siri.

Willison's read is cautious. He notes the design looks "feasible with today's technology" because vision LLMs are far better than in 2024, when Apple's first AI promises mostly did not ship. Public access is still gated behind developer betas with waitlists, so he is staying in the "believe it when I see it" camp.

Why it matters

If you build for iOS, Core AI plus screen-reading Siri changes the integration surface: features that used to require custom intents can now be triggered by Siri seeing the UI. If you watch the model market, Apple licensing Gemini rather than building its own frontier model or partnering with OpenAI or Anthropic locks in Google as the default LLM for hundreds of millions of devices.