Unveiling Apple’s On-Device and Server Foundation Models: a deep dive into Apple Intelligence

Apple introduced Apple Intelligence at the Worldwide Developers Conference 2024. This “personal” intelligence system, effectively an always-on assistant integrated into the device, will be fully embedded in iOS 18, iPadOS 18, and macOS Sequoia, and compatible with iPhone 15 Pro and later devices.

Apple Intelligence leverages advanced generative models specialized for everyday tasks such as writing and editing text, summarizing and prioritizing notifications, creating images for conversations, and executing in-app actions to streamline interactions across apps.

I have read and studied the published paper available at this link with the goal of better understanding and clarifying how this approach works.

Foundation Models

To begin with, let’s discuss the models utilized. Apple Intelligence relies on two main models: an on-device language model with approximately 3 billion parameters and a larger server-based model.

Both models are designed to offer high performance, segregate certain functions to the device and…