We introduce MJ-Bench-Video, a large-scale video preference dataset for comprehensively evaluating the reward models of text-to-video generation as well as a MoE-based video reward model