WorldCast: Distributed Multiplayer World Models

Ziyang Ye¹, Junchao Huang^1,2, Evelyn Zhang², Zhihao Xie¹, Ruicheng Zhang³, Boyao Han¹, Litao Ban⁴, Ziye Wang⁴, Xinting Hu⁵, Shaoshuai Shi⁴, Zhuotao Tian², Li Jiang^1,2†

¹CUHK-Shenzhen  ²SLAI  ³Tsinghua SIGS  ⁴Voyager Research, Didi Chuxing  ⁵USTC  ^†Corresponding author

Project page | Code | Paper (coming soon)

WorldCast is a distributed multiplayer world model. Each player runs a local client, a video generator fine-tuned from Wan2.2-TI2V-5B, on its own GPU. Clients exchange only player states, which each client projects into a camera-aligned player state field, and a shared scene state of generated blocks. This repository holds the weights of the release.

Files

file	stage	steps	dtype	size	sha256	use
worldcast_4step_bf16.safetensors	4: 4-step student (distribution matching distillation)	600	bf16	10.2 GB	8737c86b94659ba08469778fc20a5e309d9b96b6e9b4bfb3cb816d538db21f17	the generator the inference code and the demo run (paper Table 3)
worldcast_stage3_ar_fp32.safetensors	3: block-causal model with scene state	5,000	fp32	20.4 GB	e510c06a4040ee401ca1dde5f04005ad1052182f3a5b33ceca7bc1c1d1306dca	training checkpoint (initialises stage 4)
worldcast_stage2s_bidirectional_fp32.safetensors	2s: bidirectional model with player state field and scene state	25,000 (20,000 + 5,000)	fp32	20.4 GB	68883cf28a38b0cde657aa5864df453a419ed35797f09b6283468402765f9c94	training checkpoint (initialises stage 3)
depth_head.safetensors	-	-	fp32	176.8 MB	ab599fcd68115142cf3b946e147f3cb89465d0b389371cc3501e1eb3101f6e76	picture depth head of the scene state (depth of each generated block)
depth_readout.safetensors	-	-	fp32	833.4 kB	6994480d726b7e958f519e835fa306b7045cfb19171ade4aebcb594a67ccd873	read-out of the depth head
fixed_prompt_umt5xxl_bf16.safetensors	-	-	bf16	4.2 MB	a4157803a2c381835b219c079b7d811b7950c3fb2c47741d4552b7bba37cbd0a	umT5-XXL embedding of the fixed prompt (the client then skips the 11 GB text encoder)
examples/	-	-	-	201.6 MB	listed in examples/manifest.json of the code	inputs and expected videos of six recorded rounds (examples/run.sh of the code)

All generator files hold EMA weights under the parameter names of the release model (WorldCastGenerator), so a strict load_state_dict works. The training checkpoints also contain the visibility probe used during training (8 tensors). The training code is released later; the inference code needs only the 4-step model, the depth head, the read-out and the prompt embedding.

Quick start

git clone https://github.com/Ziyang-Ye/WorldCast && cd WorldCast
pip install -e .
python tools/download_weights.py --out-dir weights

This fetches the four inference files above, plus the Wan2.2 VAE, tokenizer and config.json from Wan-AI/Wan2.2-TI2V-5B, and writes weights/paths.yaml. --with-training-checkpoints also fetches the two training checkpoints; --with-taehv fetches the TAEHV tiny decoder (MIT, from its upstream repository) used by the low-latency demo. Running a session, the examples and the demo: see the code repository.

License and attribution

• The WorldCast weights are released under the Apache License 2.0.
• They are fine-tuned from Wan2.2-TI2V-5B (Apache License 2.0).
• Training data: the OpenCS2 dataset (CC BY 4.0). The files under examples/ are derived from it.
• The model renders Counter-Strike 2 game content and is intended for research. Counter-Strike 2 is a trademark of Valve Corporation; this work is not affiliated with or endorsed by Valve.

Citation

@article{ye2026worldcast,
  title   = {WorldCast: Distributed Multiplayer World Models},
  author  = {Ye, Ziyang and Huang, Junchao and Zhang, Evelyn and Xie, Zhihao and Zhang, Ruicheng and Han, Boyao and Ban, Litao and Wang, Ziye and Hu, Xinting and Shi, Shaoshuai and Tian, Zhuotao and Jiang, Li},
  journal = {arXiv preprint},
  year    = {2026}
}