Emu3
BAAI's next-token-prediction model unifying text, image, and video generation and understanding
Emu3 is an open-weight multimodal model from the Beijing Academy of Artificial Intelligence, released in September 2024, built around a single idea: treat text, images, and video all as sequences of discrete tokens and train one transformer to predict the next token across all three, with no separate diffusion or CLIP-style component. BAAI released it as an 8B-parameter model in Chat and Gen variants, one tuned for vision-language understanding and one for generation. Despite the simplified single-objective design, BAAI reported that Emu3 outperformed established task-specific open models of the time, including SDXL for image generation and LLaVA-1.6 for vision-language understanding. The project is a predecessor to BAAI’s larger Emu3.5 model, released in late 2025, which scales the same next-token-prediction approach up to 34 billion parameters.