---
library_name: lerobot
pipeline_tag: robotics
license: other
license_name: openmdw1.1-license
license_link: https://openmdw.ai/license/1-1/
base_model: nvidia/Cosmos3-Edge
tags:
  - lerobot
  - cosmos3
  - robotics
  - pretrained
---

# Cosmos3-Edge

<img
  src="https://huggingface.co/datasets/HaomingSong/lerobot-documentation-images/resolve/main/lerobot/cosmos3/cosmos3_mot_architecture.png"
  alt="Mixture-of-Transformers architecture of Cosmos 3"
  width="100%"
/>

_Mixture-of-Transformers architecture of Cosmos 3 (Figure 5 of the [technical report](https://arxiv.org/abs/2606.02800)). The understanding pathway reads the text prompt, the generation pathway reads the video latents and the actions, and the two meet in a shared attention._

[NVIDIA Cosmos3-Edge](https://huggingface.co/nvidia/Cosmos3-Edge) in the LeRobot format. This is the base checkpoint for fine-tuning as a robot policy. Taking the camera images and a language instruction as input, the policy predicts the future video frames together with the next chunk of actions.

| Item | Value |
|---|---|
| Model developer | NVIDIA |
| Original model | [nvidia/Cosmos3-Edge](https://huggingface.co/nvidia/Cosmos3-Edge) |
| Policy type | `cosmos3_edge` |
| Backbone | Mixture-of-Transformers with a dense 2B transformer, 28 layers, hidden size 2048 |
| Video tokenizer | Wan2.2 VAE |
| Parameters | 4,563,688,396 (4.6B), of which 1,423,383,232 (1.4B) train by default |
| Weights | bfloat16, the timestep embedding in float32 |
| License | OpenMDW 1.1 |

## Fine-tuning

Install LeRobot from source with the Cosmos3 extras (the Cosmos3 policy is added in [huggingface/lerobot#5134](https://github.com/huggingface/lerobot/pull/5134)):

```bash
pip install -e ".[cosmos3,training]"
```

Cosmos3 Edge can be fine-tuned in full on one 80 GB GPU (about 47 GB for 32 samples per GPU with gradient checkpointing), for example on LIBERO-10:

```bash
lerobot-train \
  --policy.type=cosmos3_edge \
  --policy.pretrained_path=lerobot/Cosmos3-Edge \
  --policy.load_action_modules=false \
  --policy.gradient_checkpointing=true \
  --policy.repo_id=${HF_USER}/Cosmos3-Edge-Policy-LIBERO_10 \
  --dataset.repo_id=lerobot/LIBERO_10-Cosmos3 \
  --accelerator.mixed_precision=bf16 \
  --batch_size=32 \
  --steps=16000 \
  --output_dir=outputs/train/cosmos3_edge_libero_10
```


- `--policy.load_action_modules=false` starts the action modules from their initialization, as the native post-training does.
- The cameras, the action dims and the frame rate come from the dataset.
- `config.json` keeps `dtype: float32` for float32 master weights in training. For inference and evaluation pass `--policy.dtype=bfloat16`.

The [Cosmos3 documentation](https://huggingface.co/docs/lerobot/cosmos3) gives the full fine-tuning recipes (multi GPU, LIBERO, DROID), the dataset requirements and the evaluation.

## Fine-tuned checkpoints

| Checkpoint | Training data |
|---|---|
| [lerobot/Cosmos3-Edge-Policy-LIBERO](https://huggingface.co/lerobot/Cosmos3-Edge-Policy-LIBERO) | [lerobot/LIBERO-Cosmos3](https://huggingface.co/datasets/lerobot/LIBERO-Cosmos3) |
| [lerobot/Cosmos3-Edge-Policy-LIBERO_10](https://huggingface.co/lerobot/Cosmos3-Edge-Policy-LIBERO_10) | [lerobot/LIBERO_10-Cosmos3](https://huggingface.co/datasets/lerobot/LIBERO_10-Cosmos3) |

## Citation

```bibtex
@techreport{nvidia2026cosmos3,
  title  = {Cosmos 3: Omnimodal World Models for physical AI},
  author = {{NVIDIA Cosmos Team}},
  year   = {2026},
  url    = {https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf}
}
```

## License

The NVIDIA model materials use [OpenMDW 1.1](LICENSE). Retain the license and the notices of [NOTICE](NOTICE) when redistributing them. The [NVIDIA model card](https://huggingface.co/nvidia/Cosmos3-Edge) describes the intended use and the limitations of the original model.