New NVIDIA Cosmos 3 Open Models for Interactive World Simulation
- ID
- 43520
- Status
- summarized
- Published
- 06 Nov 2026, 3:35 AM
- Fetched
- 06 Nov 2026, 4:22 AM
- Provider
- Hugging Face Blog
- Category
- developer-ai
- Original URL
- https://huggingface.co/blog/nvidia/cosmos-interactive-world-simulation
- Source URL
- https://huggingface.co/blog/feed.xml
Summary
- Score
- 5.0
- Created
- 06 Nov 2026, 4:23 AM
- Tags
- Audience
- ai_ml_learnersdevelopers
What happened
NVIDIA published four new open-weight checkpoints on Hugging Face as part of its Cosmos 3 world-model family: Cosmos3-Nano-Transfer-Auto (multi-view driving data with varied weather/lighting/time-of-day), Cosmos3-Nano-Sim-Auto (synchronized camera + LiDAR streams), Cosmos3-Nano-Sim-Transfer (turns simulator controls plus a text prompt into photorealistic video), and Cosmos3-Nano-Sim-Bimanual (predicts video from robot or camera actions). All four are post-trained from Cosmos 3 Nano, carry 16B trainable parameters, and ship under the OpenMDW1.1 license, which permits commercial and non-commercial use. The post itself concedes the hard part is still open: measuring where these models faithfully support policy learning and evaluation.
Why it matters
If you work on robotics or AV simulation, this is a free, commercially usable starting point you can fine-tune instead of training a world model from scratch — but the post gives zero benchmark numbers, zero FPS/latency figures, and no VRAM requirements, so budget time to run your own fidelity and throughput checks before committing. Note 'Nano' still means 16B trainable parameters: plan GPU capacity accordingly, and read the OpenMDW1.1 terms before shipping anything commercial.
Discussion angle
NVIDIA calls these 'Nano' at 16B trainable parameters and publishes no eval numbers — what would you need to see (rollout fidelity, sim-to-real gap, tokens/sec) before swapping these into an existing AV or robotics sim pipeline, and is a 16B checkpoint practical on your hardware or only via hosted inference?