Apple revealed this Monday (8) the third generation of its Apple Foundation Models (AFM), which drive the next phase of Apple Intelligence. The family includes five custom models, with significant advancements in on-device and private cloud processing, in collaboration with Google and NVIDIA.
The new models power more capable Siri, smart tools in everyday apps, and multimodal features like expressive voices and advanced image editing. All while maintaining privacy as a central principle, with data processed without storage or sharing.
Device Innovations: Efficiency with Sparse Architecture
Two models run directly on the devices. AFM 3 Core is the evolution of the dense model with 3 billion parameters, with quality gains. The highlight is the AFM 3 Core Advanced, with 20 billion parameters, natively multimodal and optimized for more advanced Apple Silicon chips.
The big news is in the architecture. Instead of loading all weights into RAM (DRAM), the model uses instruction-following pruning (IFP). The complete model is stored in flash memory (NAND) and activates only 1 to 4 billion parameters per request, depending on complexity. This resolves traditional consumer hardware limitations and enables scalability without compromising latency or power consumption.
This approach differs from conventional Mixture-of-Experts (MoE), which exchange weights token for token. Here, the routing decision is made on an entire prompt basis, with shared experts always active and others loaded only when necessary.
Private cloud models and strategic partnerships
Three models operate on Private Cloud Compute (PCC), ensuring that user data is never stored or accessed by Apple or third parties:
- AFM 3 Cloud: Optimized for speed and overall performance.
- ADM 3 Cloud (Image): focused on image generation and editing, including Image Playground and Genmoji.
- AFM 3 Cloud Pro: the most complete, for complex tasks with agents and advanced reasoning.
For the Pro model, Apple has extended PCC to NVIDIA GPUs on Google Cloud, maintaining the same privacy guarantees through confidential computing. This collaboration combines Apple’s own architecture with Google technologies (based on Gemini) and NVIDIA hardware.
What changes in practice for users and developers
The third generation represents a leap forward compared to previous generations, especially in multimodality and energy efficiency. Features like more accurate dictation, expressive voices and intelligent photo editing arrive with less dependence on the cloud for common tasks.
Not all devices that support Apple Intelligence will receive the full Advanced model — it requires high-end hardware like the iPhone 17 Pro or Macs with M4 or higher. Older models will use base or cloud AFM 3 Core.
This strategy balances local performance (fast and private) with cloud capacity for more demanding cases, differentiating itself from competitors’ all-cloud approaches.
Developers gain access to these capabilities via APIs, with deep integration into operating systems.
Apple should make the new features available for developer testing soon, with a launch for users in the fall (spring in Brazil), via free iOS, iPadOS and macOS updates.
