Apple's Foundation Models represent a fundamentally different AI strategy from OpenAI, Anthropic, and Google. Rather than competing on cloud-based model size, Apple built a family of on-device models optimized for its own hardware stack — leveraging vertical integration across chip (Neural Engine), OS (iOS/macOS), developer framework (FoundationModels), and App Store. The third-generation lineup includes a 3B dense Core model for classification/tagging and a 20B sparse Core Advanced model using Instruction-Following Pruning (IFP), which routes once per prompt to activate only 1–4B parameters, achieving ~9B-class quality at 3B compute cost. The FoundationModels Swift framework offers structured generation with @Generable types, tool calling, multimodal input, and session-based conversations — all with no API key or network required. A walkthrough builds AFMChat, a minimal on-device SwiftUI chat app using XcodeGen and a Makefile-driven workflow, demonstrating the full stack running offline on iPhone 17 Pro hardware.
