Apple Unveils Third Generation Foundation Models with Novel Sparse Architecture
Key Takeaways
- ▸Apple's AFM 3 family comprises five models: two on-device (Core and Core Advanced) and three cloud-based (Cloud, Cloud Image, Cloud Pro)
- ▸Novel sparse architecture in AFM 3 Core Advanced uses only 1-4B active parameters from a 20B total model, solving the on-device memory constraint
- ▸Partnership with Google and NVIDIA extends Apple's private cloud compute to external GPUs while maintaining privacy protections
Summary
Apple has unveiled the third generation of its Foundation Models (AFM), a comprehensive family of five custom-built models developed in collaboration with Google. The new generation includes two powerful on-device models—AFM 3 Core and AFM 3 Core Advanced—alongside three server-based models running on Apple's Private Cloud Compute infrastructure. AFM 3 Core Advanced, the flagship on-device model, introduces a novel sparsely-activated architecture using Instruction-Following Pruning (IFP) that enables a 20-billion-parameter model to activate only 1-4 billion parameters per request, dramatically reducing memory footprint while maintaining capability.
The third-generation models power a redesigned Siri and enhanced intelligent tools across Apple's operating systems, with emphasis on privacy and on-device processing. Apple collaborated with Google and NVIDIA to extend Private Cloud Compute to NVIDIA GPUs in Google Cloud while maintaining strict privacy guarantees that prevent Apple from ever storing or sharing user data. The architecture innovations enable natively multimodal capabilities, including advanced image generation, expressive voices, and enhanced dictation—all integrated deeply into the OS.
- Multimodal capabilities unlock image generation, editing tools, expressive voices, and improved dictation
Editorial Opinion
Apple's third-generation Foundation Models represent a meaningful architectural advancement in on-device AI, particularly the sparse activation approach that elegantly solves real memory constraints for consumer devices. The company's steadfast commitment to privacy—with explicit guarantees around data handling—sets an important precedent in an industry increasingly scrutinized over user data practices. However, the reliance on partnership with Google for model development and external GPU infrastructure from NVIDIA suggests that even Apple cannot achieve cutting-edge model capability entirely in-house.
