We current basis language fashions developed to energy Apple Intelligence options, together with a ∼3 billion parameter mannequin designed to run effectively on units and a big server-based language mannequin designed for Personal Cloud Compute. These fashions are designed to carry out a variety of duties effectively, precisely, and responsibly. This report describes the mannequin structure, the information used to coach the mannequin, the coaching course of, how the fashions are optimized for inference, and the analysis outcomes. We spotlight our deal with Accountable AI and the way the rules are utilized all through the mannequin growth.
This paper gives technical particulars for Apple’s On-Gadget and Server Basis Fashions, launched on June 10, 2024, on this submit.