Overview
- Apple provided technical clarifications after its WWDC keynote, saying the third‑generation Apple Foundation Models (AFM) family now has five models spanning two on‑device and three cloud variants.
- The most powerful on‑device model, AFM 3 Core Advanced, is a 20‑billion‑parameter sparse model that stores the full model in flash and activates just 1–4 billion parameters per request to run on top‑tier Apple silicon such as A19 Pro.
- Apple maps AFM Cloud Pro to NVIDIA GPUs hosted in Google Cloud under its Private Cloud Compute (PCC) architecture while the other cloud models run on Apple servers or Apple silicon.
- Apple says it pre‑trained and fine‑tuned AFMs with its own data and reinforcement learning, then refined them using outputs from Google’s Gemini frontier models rather than adopting Google’s deployed Gemini instances or client code.
- To protect data, Apple describes PCC attestation patterns, NVIDIA confidential compute, a cryptographic, append‑only hardware ledger and researcher access through its security program to let outside experts verify the company’s privacy claims.