Based on their presentations yesterday and accompanying written materials they have a pretty capable ~3B parameter Foundational LLM running fully locally on the devices which is the first line. https://machinelearning.apple.com/research/introducing-apple...