“An 8-member ensemble model with 2.4 billion total parameters can be distilled into a single 300 million parameter model while retaining 83% of the loss improvement.”