They probably just beefed up compute run time on the what is the same underlying model