SolarWM / WORLD MODELING

SolarWM Open Data and Scalable Training for Long-Horizon Video World Models

Junchao Huang1,2Guian Fang3Shengju Qian4Xianghao Kong5Zhuoran Zhao5,6

Wei Huang7Yihua Du6Zixin Zhang6Justin Cui8Yuchao Gu7Yukang Chen7Xinting Hu

Tianyu He9Shaoshuai ShiZhuotao Tian2Xin WangMike Zheng Shou3Li Jiang1,2

1CUHK-SZ2SLAI3NUS4CUHK5HKUST6HKUST-GZ7NVIDIA8UCLA9MSRA

Paper Code Checkpoint Dataset Dataset Dataset Access Form

OPEN INFRASTRUCTURE · 2026

01 / OVERVIEW

Sound is on by default. Use the player controls to adjust the volume.

KEY FEATURES

EFFICIENT REAL-TIME INFERENCE

Imagine worlds.
Interact in real time.

Trained only on five-second clips, the model enables minute-scale and even hour-scale real-time inference while preserving responsive interaction and coherent world evolution over extended rollouts.

OPEN DATA PIPELINE

1.43M clips · 25 TB of data.
Fully open.

We release the complete data-processing stack: scale-consistent camera annotations, fine-grained filtering, and multiple training recipes.
Explore the dataset

ONE FRAMEWORK, MANY BACKBONES

1 framework.
4 base models.

A unified framework supports state-of-the-art video models across architectures and parameter scales, including SolarWM-Wan-5B, SolarWM-Wan-14B, SolarWM-LTX-2.5, and SolarWM-Minimax-H3.
02 / DEMOS

Generated by the SolarWM-Wan-5B Causal Student (4 steps)

06 / CITATION 
@misc{huang2026solarwmopendatascalable,
  title={SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models},
  author={Junchao Huang and Guian Fang and Shengju Qian and Xianghao Kong and Zhuoran Zhao and Wei Huang and Yihua Du and Zixin Zhang and Justin Cui and Yuchao Gu and Yukang Chen and Xinting Hu and Tianyu He and Shaoshuai Shi and Zhuotao Tian and Xin Wang and Mike Zheng Shou and Li Jiang},
  year={2026},
  eprint={2609.02886},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.02886},
}