Apple wants to train AI on your private personal data
Apple introduced its third generation of Apple Foundation Models (AFM 3), describing a family of five foundation models custom-built in collaboration with Google that spans on-device models to server-based models running on Private Cloud Compute. Apple says the architecture puts privacy at its core and is designed to unlock experiences such as an entirely new Siri and intelligent tools embedded across everyday apps.
The on-device lineup includes two models: AFM 3 Core, the next generation of Apple's 3-billion-parameter dense model that delivers a step up in quality, and AFM 3 Core Advanced, its most powerful on-device model. AFM 3 Core Advanced is natively multimodal, enabling features like expressive voices and higher-accuracy dictation, and uses a 20-billion-parameter sparse architecture that activates only 1 to 4 billion parameters at a time depending on the request. It is unlocked by and optimized for Apple's most capable Apple silicon systems.
The three server-based models run on Private Cloud Compute, which Apple says ensures user data is never stored or shared with anyone, including Apple. They are AFM 3 Cloud, a server-side workhorse optimized for speed, efficiency, and performance; ADM 3 Cloud (Image) for image generation and editing, which powers advanced photo-editing tools and the all-new Image Playground; and AFM 3 Cloud Pro, the most capable server-based model, aimed at demanding use cases like agentic tool use and complex reasoning. AFM 3 Core, AFM 3 Core Advanced, AFM 3 Cloud, and ADM 3 Cloud are all purpose-built for Apple silicon, while for AFM 3 Cloud Pro Apple worked with Google and NVIDIA to extend Private Cloud Compute to NVIDIA GPUs in Google Cloud while maintaining the same privacy guarantees. Apple points to its Security Research website for more details.
Apple says the third generation delivers significant advancements in capabilities and quality, and that its overview explores the scalable architectures and training methodologies behind the on-device and server models, which are integrated deep into its operating systems. It highlights AFM 3 Core Advanced as an area of deep innovation: traditional large language models, whether dense or sparsely activated, require all weights to reside in active memory (DRAM), creating a massive footprint that limits scalability on consumer hardware — a barrier the post sets out to explain how Apple broke (the excerpt cuts off mid-sentence at that point). The Hacker News submission surfaced the news under the headline 'Apple wants to train AI on your private personal data,' framing the announcement around the privacy trade-offs of personalization.