Gamma-World’s 24fps responsiveness plus four-player collaboration—are game engines about to get shaken up? With Tsinghua, Toronto, and Vector’s “top three” teaming up, this one really has a lot going for it.

NVIDIA released the multi-agent generative world model Gamma-World, supporting four-player collaboration and 24fps response, developed jointly with Tsinghua University, the University of Toronto, and the Vector Institute. The model introduces rotational position encoding and information intermediary markers, achieving zero-shot generalization from two to four players for the first time without retraining. To handle increased computational load as the number of players grows, sparse center attention is used, making the cost linear. Evaluations show that it outperforms traditional networks in terms of visual realism, action controllability, and player consistency.