[jjfoooo4]

My guess is that they will lean towards smaller models, and try to provide the best experience for running inference on device