In this paper, we present a new model, called LongT5, with which we
explore the effects of scaling both the input length and model size at the same time.