Apple Launches Five Third-Gen Foundation Models, Co-Built With Google and NVIDIA
Summary: Apple at WWDC 2026 unveiled its third-generation Apple Foundation Models (AFM 3), a five-model family co-developed with Google and NVIDIA that pushes on-device intelligence to 20 billion parameters while extending Private Cloud Compute to Google Cloud infrastructure.
Key Facts
- AFM 3 Core (3B dense): next-gen on-device model with improved quality across Apple Intelligence tasks
- AFM 3 Core Advanced (20B sparse multimodal): Apple's most powerful on-device model to date — sparse architecture activates only 1–4B parameters per request, keeping it viable on Apple silicon
- AFM 3 Cloud / ADM 3 Cloud / AFM 3 Cloud Pro: server-side models on Private Cloud Compute, with Cloud Pro running on NVIDIA GPUs in Google Cloud under the same privacy guarantees
- The release is the first tangible output of the Apple–Google multi-year partnership announced in January 2026
Why It Matters
Apple has long staked its AI story on on-device privacy. Openly leveraging Google Cloud for its highest-tier model is a meaningful shift. The 20B sparse on-device model raises the bar for what's expected from smartphone-class AI, and sets up a direct benchmark against competing on-device approaches from Google and Samsung.
Further Reading
- Apple ML Research blog — Apple
- Full breakdown — Let's Data Science