Tencent opens Multics, a world model for 10 players

The University of Hong Kong and Tencent's ARC Lab released Multics, an open multi-player world model that generates video of one shared scene from several players' viewpoints. The code and the paper went up on 12 October 2026, and the weights sit on Hugging Face.

Multics organizes communication across three levels: player, group, and global. Group tokens enter the self-attention as intermediate representations that gather locally relevant player information, then connect to global tokens for cross-group communication. Because the group-token parameters are shared across group slots, the same module carries different numbers of players, and a model trained with 4 players generalizes to 6, 8, and 10 players without retraining, according to the repo and the paper.

Given each player's first frames, a text prompt, actions, and a camera trajectory, Multics jointly generates their videos and keeps overlapping views consistent. The demo config targets 4 players and 165 frames at 50 inference steps, and the repo includes prepared examples for 6, 8, and 10 players. The paper's abstract reports state-of-the-art performance in the 4- and 6-player settings on cross-player consistency, action and camera controllability, and robustness to interference between unrelated events, without figures.

Grouping is driven by recent camera trajectories, which limits how much unrelated players communicate. The paper says the hierarchy cuts self-attention computation against fully connected approaches such as Solaris, whose cost grows quadratically with players, and avoids the interference that centralized approaches such as gamma-World see in a shared global representation. Multics is built on Wan and DiffSynth-Studio.

Multics is research video generation. It does not run as a real-time playable game at frame rates a player controls. Running it needs Linux, Python 3.12, an NVIDIA GPU, the CUDA 12.8 toolkit, and a C++ compiler. The Hugging Face bundle holds the 4-player checkpoint, the Wan2.2 VAE, the UMT5 text encoder, and the tokenizer.

The authors are Kaiyi Huang, Wangbo Yu, Yukun Huang, Jiwen Yu, Kunhao Liu, Donglin Yang, Wenbo Hu, Ying Shan, and Xihui Liu, from The University of Hong Kong and ARC Lab, Tencent IEG. Wangbo Yu and Xihui Liu are the corresponding authors.

For creators, the useful claim is the scalability: one trained checkpoint that holds as player counts grow is the piece open multi-view AI video has been missing, if the outputs hold up outside the authors' own examples. The clips on the project page are the only published ones to judge that against.

Get it from the repo with hf download TencentARC/Multics --local-dir models, then run inference.py; it needs a Linux machine with an NVIDIA GPU.