Apple has announced a new generation of Apple Foundation Models (AFM), featuring two on-device models and three cloud-based variants. The largest on-device model, AFM 3 Core Advanced, contains 20 billion parameters and supports multimodal tasks.
Apple's latest foundation models represent the next iteration of its Apple Intelligence initiative, designed to process AI tasks directly on devices while maintaining user privacy. The on-device models enable local processing of machine learning tasks without relying on cloud infrastructure.
The 20-billion parameter AFM 3 Core Advanced handles multimodal inputs and outputs, allowing it to work across text, images, and other data types. This scale positions it as a capable alternative to larger cloud-dependent models while avoiding network latency and external data transmission.
Beyond the on-device offerings, Apple has developed three cloud-based foundation models to handle more computationally intensive tasks. This hybrid approach—combining on-device and cloud capabilities—allows Apple to balance user privacy with the computational demands of advanced AI features.
The models are deeply integrated into Apple's operating systems, suggesting broader deployment across iOS, iPadOS, macOS, and other platforms. This integration enables AI features to function seamlessly within existing Apple applications and services.
Apple's foundation model strategy reflects the industry shift toward deploying AI across multiple tiers. On-device processing reduces dependence on cloud infrastructure, cuts latency, and limits data exposure. Cloud models provide additional capacity for users who opt into server-based processing.
The announcement comes as Apple faces regulatory scrutiny in Europe over cloud service interoperability and continues efforts to address smartphone theft through device security measures. The foundation models may support future features tied to these initiatives, though Apple has not detailed specific applications.
Apple Machine Learning Research shared the announcement, suggesting the models may support both commercial deployment and research applications.
Researchers are using AI to decode animal communication, offering unprecedented insight into other species. Bioethicists warn the technology could enable new forms of manipulation and exploitation.
Artificial intelligence-generated attack ads are flooding social media ahead of Victoria's election, with mysterious sponsors outspending major political parties. The ads feature alarming scenarios—machete attacks, hospital crises—that echo opposition talking points.
A group of hikers required emergency rescue after relying on Google's Gemini AI for trip planning, which allegedly recommended dangerously insufficient food and water supplies.
OpenAI has published a comprehensive prompting guide for GPT-6 Astra that helps developers optimize interactions with the model. The guide includes techniques for increasing model initiative and a blocklist of common AI "slop" phrases to avoid.