Emu3 is a state-of-the-art multimodal model trained solely through next-token prediction, capable of handling images, text, and videos. It surpasses several flagship models on generation and perception tasks without the need for diffusion or compositional architecture. By unifying multimodal sequences into a single transformer model, Emu3 simplifies the complexity of multimodal model design, demonstrating significant potential for scaling during both training and inference.