pi-H3 is a distilled version of the MiniMax H3 model that produces videos with audio in two or four steps. It predicts final pixels directly from the last sampling step instead of using a separate decoder.