Apple revealed its third‑generation Foundation Models at WWDC26, expanding the lineup to five distinct systems. Two stay on device—AFM 3 Core and AFM 3 Core Advanced—while three run on servers, including AFM 3 Cloud and the cloud‑only AFM 3 Cloud Pro that leverages NVIDIA GPUs in Google Cloud. The mix aims to balance speed, privacy and power for consumer experiences and developer toolkits today and tomorrow.

AFM 3 Core Advanced pushes the on‑device envelope with 20‑billion parameters, yet activates only 1‑4 billion at a time through a sparse, Apple‑invented mechanism. This keeps the model lightweight while delivering expressive voices and high‑accuracy dictation. The server‑side AFM 3 Cloud focuses on speed and efficiency for text generation across Apple silicon for developers and end users who require instant responses.

The cloud‑only AFM 3 Cloud Pro runs on NVIDIA GPUs hosted by Google, a first for Apple after extending its Private Cloud Compute to third‑party infrastructure. Apple claims the partnership preserves its privacy guarantees through a cryptographically verifiable ledger and multi‑root attestation. The model powers agentic tool use and complex reasoning tasks for enterprise applications and consumer AI.

Human evaluations show AFM 3 Core and AFM 3 Cloud outperform previous generations in instruction following, truthfulness, and image understanding across multiple locales. AFM 3 Core Advanced also beats Apple’s existing dictation system on formatting and comprehension. With a unified foundation, Apple signals a shift toward multimodal AI that balances on‑device privacy with cloud scale for today’s data‑driven world and future innovation.