Apple's description directly rejects the claim that it trains the models on customers' private personal information. The company says its training mixture includes public material, licensed or purchased datasets, information from dedicated studies, open-source data and synthetic examples. It says private user data and user interactions are not used for foundation-model training, and that web publishers can opt out of having their material included.
Two models are designed for on-device use and three run through Private Cloud Compute. Apple says that cloud environment does not store or share user data, including with Apple. Most models are tailored to Apple silicon, while AFM 3 Cloud Pro is optimized for Nvidia graphics processors hosted in Google Cloud under an extension of the Private Cloud Compute system.
The most technically distinctive device model is AFM 3 Core Advanced. Large models ordinarily need their active weights in DRAM, limiting how much can fit on consumer hardware. Apple's proposed solution stores the full model in flash memory and selects a smaller collection of expert weights for each prompt. A compact dense component makes the initial routing choice and can periodically select new experts during generation. Shared experts remain active, while other experts move into memory only when needed.
Apple says this prompt-level routing is designed to reduce slow transfers between flash storage and DRAM. It also allows different requests to activate different predetermined amounts of the model, potentially matching computational effort to task difficulty without maintaining several separate small models. The practical balance among latency, energy use and quality will depend on the devices and features in which Apple deploys it.
For server work, AFM 3 Cloud expands a mixture-of-experts architecture intended for multimodal reasoning and long-context recall. A separate image model, ADM 3 Cloud, supports image generation, editing and Genmoji, with adapters for features including spatial reframing and touch-based changes.
Apple says the family began from a common foundation before specialization and was post-trained with supervised fine-tuning and multiple stages of reinforcement learning. The announcement presents a broad architectural plan, but real-world assessments will need to test individual features. Its clearest privacy statement is also narrower than some online characterizations: Apple explicitly says private personal data and interactions are excluded from foundation-model training.
