SolarWM is an open foundation for building interactive video world models from data preparation through scalable training and long-horizon inference. This repository contains only the MiniMax-H3 33B family.

Checkpoints

Recommended for inference: SolarWM-h3-33B-sgf-stage2-158f-fix uses the step 3900 EMA weights after longer DMD training than the original step 1200 release. The additional DMD training reduces blur in generated videos, particularly in trees. The original step 1200 EMA checkpoint remains available for reproducibility.

Download

Accept the repository's access terms, then authenticate:

python -m pip install --upgrade huggingface_hub
hf auth login
export SOLAR_MODEL_ROOT=/path/to/SolarWM-models

For inference, download the recommended Stage2 fix checkpoint below. It is the only stage checkpoint you need. The base model and runtime assets are also required; skip the base download if already installed.

hf download junchaoh-cs/SolarWM-H3-33B \
  --include "SolarWM-h3-33B-base/**" "SolarWM-h3-33B-sgf-stage2-158f-fix/**" \
  --local-dir "$SOLAR_MODEL_ROOT"

To download only the recommended Stage2 fix checkpoint into an existing H3 setup:

hf download junchaoh-cs/SolarWM-H3-33B \
  --include "SolarWM-h3-33B-sgf-stage2-158f-fix/**" \
  --local-dir "$SOLAR_MODEL_ROOT"

For training across stages, download the full repository, including the base model and all EMA checkpoints:

hf download junchaoh-cs/SolarWM-H3-33B --local-dir "$SOLAR_MODEL_ROOT"

Keep each checkpoint directory intact and select checkpoint.weight_source=ema. When switching from the original Stage2 release, update your inference checkpoint path to the downloaded SolarWM-h3-33B-sgf-stage2-158f-fix directory. These packages contain weights and loading manifests; they do not contain the optimizer or data-reader state needed to resume an interrupted training run.

Inference

Install SolarWM from the source repository, activate the MiniMax-H3 runtime environment, and follow the MiniMax-H3 guide.

License

The MiniMax-H3 packages in this repository are governed by the MiniMax-H3 Community License included in the release directories, including its territory restrictions. They are not relicensed under the Apache-2.0 license used by SolarWM code. Review the applicable license before use or redistribution.

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}
}