The training uses Apple’s AXLearn (which runs on TPUs and Apple Silicon), Server model inference runs on Apple Silicon (! Apple has considerable published work on image models, more so than language models (compare the amount of each model type on Apple’s HF page). The on-device models will come with a set of LoRAs and/or DoRAs (Adapters, in Apple parlance) that specialize the on-device model to be very good at specific tasks.