Multics

Multics is a generative multi-player world model from researchers at The University of Hong Kong and ARC Lab, Tencent IEG. It simulates a shared environment from several players' perspectives at once. Given each player's initial view, a text prompt, actions and a camera trajectory, it jointly generates one video per player. Overlapping views stay consistent and each player follows their own controls.

The method uses hierarchical communication across player, group and global levels. Group tokens are added to self-attention as intermediate representations, and their parameters are shared across group slots, so the same module handles different player counts. Dynamic grouping based on recent camera trajectories assigns relevant players to groups under a fixed budget. The authors compare it against fully-connected (Solaris) and centralized (γ-World) approaches.

The project page reports state-of-the-art results in 4- and 6-player settings. The model was trained with four players and run with 6, 8 and 10 without retraining. At 10 players it cuts DiT latency by 61.9%, self-attention latency by 78.1% and self-attention FLOPs by 86.1% compared with fully-connected communication. A YouTube creator says code has been released, but no repository link appears in the provided material. The page's six-player comparison section is marked as not yet available. The work is a research project, not a hosted product.