Apple unveiled its third‑generation Apple Foundation Models (AFM) during a post‑keynote tech talk. Senior VP Craig Federighi and AI leaders outlined that the new models rely entirely on Apple’s own architecture, using no Google Gemini models, client‑side code, or Google Search infrastructure. The company stresses full independence from Google’s AI stack.

Subramanya mapped the AFM family to two on‑device and three server‑side models. On‑device, AFM Core uses a dense architecture, while AFM Core Advanced adopts a sparse, multimodal design that enables invitation and expressive‑voice features without cloud calls. Server models—AFM Cloud, AFM Cloud Image, and AFM Cloud Pro—handle latency‑optimized compute, image generation, and complex reasoning.

Training followed a distillation approach: models were built for Apple Silicon, trained on proprietary data with reinforcement learning, and refined using outputs from Gemini frontier models. To power AFM Cloud Pro, Apple partnered with Google and Nvidia, extending its private cloud to Nvidia GPUs in Google’s cloud while keeping data invisible through Nvidia’s ambiguous confidential compute.

Federighi highlighted a System Orchestrator that routes queries to the right model—on‑device or cloud—based on complexity and context. The architecture taps an App Toolbox, Spotlight Semantic Index, and a World Knowledge Service built over years. Apple assures that its Private Cloud Compute, including the Nvidia‑GPU extension, can be audited to confirm no user data is stored.