The Beijing Academy of Artificial Intelligence (BAAI) has launched Emu3, a state-of-the-art multimodal model trained via next-token prediction. Emu3 is capable of generating images by predicting vision tokens and supports various resolutions and styles. It outperforms existing models such as sdxl in image generation, llava 1-6 in image understanding, and open sora in video generation. Emu3 features 9 billion parameters and includes a state-of-the-art tokenizer. The model is available with open weights and can be accessed on Hugging Face. A demo is also available.