we-dream/pretrain-latent-2kh

WE-Dream latent-action pretraining — 2kh

WE-Dream is a humanoid world-action model (WAM). This checkpoint was pretrained on 2,000 hours of egocentric video with 32-dimensional latent actions and saved at training step 50,000. 2kh refers to the pretraining data scale, not robot demonstration duration.

The package includes sharded safetensors weights, model configuration, transforms, and normalization metadata for robot post-training. The action horizon is 24 and the configured routed-attention probability is 0.5. This is a pretraining checkpoint; it is not a ready-to-deploy G1 controller.

Usage

Use the WE-Dream codebase and its checkpoint instructions. The checkpoint requires the project's custom model classes and dependency setup.

from huggingface_hub import snapshot_download

checkpoint_dir = snapshot_download(
    repo_id="we-dream/pretrain-latent-2kh",
    local_dir="checkpoints/we-dream-latent-2kh",
)

If prompted, request access on this model page and log in with hf auth login. For reproducible experiments, set revision to the release commit you downloaded.

Source and license

The model weights and runtime metadata are copied without conversion from the released ModelScope checkpoint at revision 049654c08c8c4c41eaec29ca9b404a247de1d504. See provenance.json for source sizes and SHA-256 hashes. The source checkpoint declares the Apache-2.0 license; this mirror preserves that declaration.