Apple Reveals 5-Model Siri AI Stack Built on Gemini, Keeping 2 Models On-Device
Updated
Updated · 9to5Mac · Jun 11
Apple Reveals 5-Model Siri AI Stack Built on Gemini, Keeping 2 Models On-Device
3 articles · Updated · 9to5Mac · Jun 11
Summary
Apple’s Siri AI uses a five-model architecture built with Google: four models are customized Gemini variants tuned for Apple Silicon, while the largest runs on dedicated Google servers.
Two of those models run entirely on-device, and two more run on Apple’s Private Cloud Compute servers, with Apple saying no user data is retained or exposed to either Apple or Google.
The company says the Apple Silicon models were trained on proprietary data, refined with outputs from Gemini frontier models, and use Apple’s own knowledge sources rather than Google Search or Google’s knowledge graph.
For the most powerful model on Google servers, Apple says the same PCC protections apply — stateless processing, no privileged runtime access, non-targetability and verifiable transparency — though that setup remains less tested than Apple’s own infrastructure.