The Third Generation of Apple's Foundation Models
7 hours ago
- #p
- #A
- #d
- #g
- #e
- #r
- #a
- #u
- #c
- #P
- #t
- #o
- #v
- #M
- #i
- #F
- #
- #,
- #y
- #s
- #I
- #n
- #l
- Apple introduced its third generation of Apple Foundation Models (AFM) to power Apple Intelligence, with a focus on user-centricity, deep OS integration, and privacy-first architecture.
- The AFM family includes five models: two on-device (AFM 3 Core and AFM 3 Core Advanced) and three server-based (AFM 3 Cloud, ADM 3 Cloud for images, and AFM 3 Cloud Pro), all optimized for Apple silicon except AFM 3 Cloud Pro, which uses NVIDIA GPUs via Google Cloud.
- AFM 3 Core Advanced features a novel sparse architecture using Instruction-Following Pruning (IFP) to enable on-device multimodal capabilities with efficient memory usage, activating only 1-4 billion parameters per request.
- Server models like AFM 3 Cloud utilize Paralell-Track Mixture-of-Experts (PT-MoE) enhancements for improved reasoning and recall, while ADM 3 Cloud enables advanced image generation, editing, and Genmoji.
- Training involved diverse high-quality data, advanced pre-training on cloud TPUs, and post-processing with supervised fine-tuning plus reinforcement learning, all without using user private data.
- Evaluations show significant improvements over previous generations in text and image tasks, with human graders noting better instruction following, truthfulness, and speech quality in TTS and dictation.
- Apple emphasizes Responsible AI principles, including user empowerment, bias mitigation, safety, and privacy protection through on-device processing and Private Cloud Compute.
- The new models enable an enhanced Siri, expressive voices, advanced photo editing, and Image Playground features, all running securely on-device or via Private Cloud Compute.