MiniWAM

Checkpoints for MiniWAM: Learning Compact Future Targets for Efficient World-Action Modeling (paper, code, project page).

MiniWAM is a 0.25B world-action model. Instead of predicting future video or native visual features, it predicts a compact future target, PRISM, alongside the action chunk. Both checkpoints use frozen DINOv3 ViT-B/16 features and UMT5-XXL instruction embeddings; the parameter count excludes these encoders.

Folder	Contents	Results
libero/	LIBERO checkpoint and its instruction embeddings (libero_instruction_features.pt)	LIBERO 97.2%, LIBERO-Plus (10,030 tasks) 73.6%
robotwin/	RoboTwin 2.0 checkpoint	46.1% (clean 73.4%, randomized 18.8%)
libero_plus_text_features/	Instruction embeddings for evaluating libero/ on LIBERO-Plus: all 10,030 tasks and the 2K ablation subset
pretrained/	Third-party frozen encoders, unmodified (see Attribution)

Each checkpoint folder holds model.safetensors, config.yaml, statistics.json and metadata.json.

Download

From the root of the MiniWAM repository:

hf download J1dan/MiniWAM --include 'libero/*' 'robotwin/*' --local-dir checkpoints
hf download J1dan/MiniWAM --include 'pretrained/*' --local-dir .   # UMT5-XXL encoder, Wan2.1 VAE
hf download J1dan/MiniWAM --include 'libero_plus_text_features/*' --local-dir datasets

LIBERO evaluation needs only the first command, and LIBERO-Plus evaluation also needs the third. The encoders in pretrained/ (12 GB) are needed only to extract training features, to build instruction caches, for the official RoboTwin protocol and for the Wan-VAE ablations. RoboTwin evaluation replays the frozen cases in J1dan/MiniWAM-RoboTwin-frozen-cases.

DINOv3 ViT-B/16, the frozen visual encoder, is not included. Request access at facebook/dinov3-vitb16-pretrain-lvd1689m (DINOv3 License), then download it once; MiniWAM loads it from the Hugging Face cache:

hf auth login
hf download facebook/dinov3-vitb16-pretrain-lvd1689m \
  --revision 5931719e67bbdb9737e363e781fb0c67687896bc

SHA-256 checksums of the large files, by their path in this repository:

b8edd07be8fe26e358b9f9ee8d291f1f432d046d748c8bbf92741d85178f7f22  libero/model.safetensors
485c925a8291d6fd255262e687fae7f61b81d88cc7879d0c9a9012e1ed50a336  robotwin/model.safetensors
9c5467e1cfa69a7e2dcb747ae569549679d640b75c76108cc8a094e8c0941a59  libero_plus_text_features/libero_plus_full_10030.pt
c67048f85bf030a42060bd03c017fe09ec9e82ff61b1c0eb0e65ccd72be5e7d9  libero_plus_text_features/libero_plus_2k.pt
7cace0da2b446bbbbc57d031ab6cf163a3d59b366da94e5afe36745b746fd81d  pretrained/language/models_t5_umt5-xxl-enc-bf16.pth
38071ab59bd94681c686fa51d75a1968f64e470262043be31f7a094e442fd981  pretrained/vision/Wan2.1_VAE.pth

Attribution

The models in pretrained/ were not made by the MiniWAM authors. They are byte-for-byte copies of the upstream files at the revisions below, kept here so that the exact files behind the MiniWAM results stay available. Each keeps its license, included next to it. Please cite the original work.

File	Original work and authors	Copied from (repository @ revision)	License
pretrained/language/models_t5_umt5-xxl-enc-bf16.pth	UMT5-XXL text encoder. Google (Chung et al., 2023); encoder-only bf16 export by the Wan Team, Alibaba	Wan-AI/Wan2.1-T2V-14B @ b62153b54fce61f5908de23f6d8715e147de0daf	Apache-2.0
pretrained/vision/Wan2.1_VAE.pth	Wan2.1 video VAE. Wan Team, Alibaba (2025)	Wan-AI/Wan2.1-VACE-14B @ ab1867ace0fb7738d46c46aab8308b71c51bb011	Apache-2.0

Neither upstream repository includes a NOTICE file. The VAE is used only by the Wan-VAE ablations.

The instruction embeddings (libero/libero_instruction_features.pt and libero_plus_text_features/) are UMT5-XXL outputs for instructions from LIBERO (MIT) and LIBERO-Plus. The LIBERO checkpoint was trained on LIBERO-fastwam (CC BY 4.0, derived from LIBERO), and the RoboTwin checkpoint on RoboTwin 2.0 demonstrations (RoboTwin, MIT).

License

The MiniWAM checkpoints and instruction embeddings are released under the MIT license. The files in pretrained/ are under Apache-2.0. See LICENSES.md.

Citation

@article{chen2026miniwam,
  title   = {MiniWAM: Learning Compact Future Targets for Efficient World-Action Modeling},
  author  = {Chen, Jie and Bai, Ruofei and Cai, Yuxin and Zhang, Yifeng and He, Chengyang and
             Li, Jun and Yau, Wei-Yun and Sartoretti, Guillaume},
  journal = {arXiv preprint arXiv:2610.12194},
  year    = {2026}
}

If you use the encoders in pretrained/, please also cite their authors:

@article{wan2025wan,
  title   = {Wan: Open and Advanced Large-Scale Video Generative Models},
  author  = {{Wan Team} and Wang, Ang and Ai, Baole and Wen, Bin and Mao, Chaojie and
             Xie, Chen-Wei and Chen, Di and Yu, Feiwu and Zhao, Haiming and Yang, Jianxiao and others},
  journal = {arXiv preprint arXiv:2503.20314},
  year    = {2025}
}

@inproceedings{chung2023unimax,
  title     = {UniMax: Fairer and More Effective Language Sampling for Large-Scale Multilingual Pretraining},
  author    = {Chung, Hyung Won and Garcia, Xavier and Roberts, Adam and Tay, Yi and
               Firat, Orhan and Narang, Sharan and Constant, Noah},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2023}
}