Emu3.5: Ushering in a New Era of Multimodal World Models

The Beijing Academy of Artificial Intelligence (BAAI) has released "Emu3.5," ushering in a new era of large-scale multimodal world models. Emu3.5 is a dense autoregressive Transformer model with 34 billion parameters, which unifies the modeling of multimodal data such as images, text, and videos, and achieves a leap in capabilities from "next token prediction" to "next state prediction."