Multix brings multiplayer gaming to generative world models

Multix (named Multics in the repository) is a generative multi-player world model developed by Tencent's ARC Lab and The University of Hong Kong. It simulates multiple players interacting inside the same virtual environment, generating distinct camera perspectives for each player while keeping all character actions and surroundings fully in sync.

Most existing world simulators, like Genie 3 or Hunyuan World, are limited to single-player viewpoints. When multiple actors enter the frame, consistency usually falls apart. Multix directly tackles this multi-agent synchronization challenge.

How it scales to 10 players

Simulating several interacting players simultaneously usually leads to an exponential explosion in attention computation. Multix gets around this with a dynamic architecture called hierarchical communication.

Instead of forcing every player's visual state to attend to everyone else at every step, the model organizes information flow across three tiers: player level, group level, and global level. It clusters players dynamically based on their current 3D positions and camera angles, only sharing context between players who actually influence one another. This design cuts attention computation by more than 86%.

Even better is its generalization trick. The researchers trained the checkpoint using data from just 4 players. Thanks to shared group-token parameters, the exact same model scales up to handle 6, 8, or 10 players seamlessly, without any fine-tuning or retraining.

Getting it running locally

The code is released under the Apache-2.0 license on GitHub, accompanied by checkpoints on Hugging Face, inference scripts, and full training pipelines.

The base inference bundle weighs in at around 16GB. The framework is built on top of the Wan2.2 video backbone (specifically Wan2.2-TI2V-5B) and DiffSynth-Studio, taking advantage of the open OpenCS2 dataset.

To run it on your own rig, you will need a Linux machine equipped with an NVIDIA GPU, Python 3.12, CUDA 12.8, and a C++ compiler. Under BF16 on a single GPU, generating a demo outputs 640 × 352 resolution per player across 165 frames over 50 diffusion steps. VRAM consumption scales roughly linearly as you ramp up from 4 to 10 players.

Why it matters

We think Multix shifts generative world models from passive solo video clips toward actual multi-agent game engines. Game developers and embodied AI researchers should definitely pull the repository: the hierarchical token-sharing pattern offers a practical blueprint for training interactive multi-character simulations without burning infinite compute.

Keep in mind that output resolution is still capped at 640 × 352, and diffusion latency means you cannot play it at native 60 FPS just yet. But proving that a 4-player trained weight can generalize to a 10-player room while slashing compute by 86% is a serious architectural breakthrough.