Apple’s third-gen Foundation Models explained

At this year’s WWDC26 keynote, Apple unveiled the third generation of its Apple Foundation Models (AFM). These updated AI systems are designed to power a range of on-device and cloud-based features, from Siri improvements to advanced photo editing. Here’s a clear look at how each model functions and what sets them apart.

The latest AFM lineup includes three distinct models: a small on-device model, a larger server-based model, and a specialized language model for specific tasks. The on-device variant, optimized for privacy and speed, runs directly on iPhones and iPads using the Neural Engine. It handles real-time requests like text prediction and voice commands without sending data to the cloud, reducing latency and protecting user information.

Meanwhile, the larger server model operates on Apple’s private cloud infrastructure. It tackles more complex queries, such as summarizing long documents or generating creative content, while still adhering to Apple’s strict privacy standards. This model uses a transformer architecture with enhanced attention mechanisms to improve context understanding and response accuracy.

The third model focuses on multimodal capabilities, integrating text, images, and audio inputs. For instance, it can analyze a photo and answer questions about its content, or generate captions based on visual cues. Apple trained this model on a diverse dataset to ensure it handles varied tasks reliably, from accessibility features to creative tools in apps like Photos and Notes.

All three models share a common foundation: they are built on Apple’s proprietary chip architecture, which allows for efficient processing and low power consumption. Training involved a mix of supervised learning and reinforcement learning from human feedback, fine-tuning outputs to align with user expectations. Apple also emphasized that these models are designed to minimize bias and maintain factual accuracy, with regular updates based on real-world usage.

In practice, the AFM third generation improves Siri’s ability to handle multi-step requests, enhances autocorrect and predictive text, and powers new features like real-time language translation. Developers can also access these models through Core ML APIs to integrate AI into their apps, offering customization while keeping data on-device whenever possible.

Apple’s approach prioritizes privacy and performance over raw scale. Unlike competitors that rely on massive cloud models, Apple balances local processing with server assistance, ensuring that sensitive data stays protected. The result is a suite of AI tools that feel responsive and secure, setting a new standard for consumer-facing artificial intelligence.