Lots and lots of synthetic data from the bigger models training the smaller ones would be my guess.