NVIDIA released four new open-weight Cosmos 3 Nano checkpoints on November 5, 2026, covering interactive driving simulation, synchronized camera-and-LiDAR streaming, simulator-to-video rendering, and two-arm robot prediction. Why it matters: physical-AI teams now get downloadable, commercially usable world models for the two hardest data problems in robotics and self-driving — generating believable sensor data and closing the loop between actions and observations.
November 5, 2026
What happened
NVIDIA expanded its Cosmos world-model platform with four post-trained checkpoints, each built from Cosmos 3 Nano with 16 billion trainable parameters, published on Hugging Face under the OpenMDW 1.1 license, which permits commercial and non-commercial use under its terms. NVIDIA describes them as post-trained examples to build on — not finished self-driving or robotics products.
The four checkpoints
Cosmos3-Nano-Transfer-Auto generates synchronized multi-view driving data — up to 11 camera views, including fisheyes, plus joint RGB and LiDAR — with varied weather, lighting, and time of day, for offline synthetic data generation. Cosmos3-Nano-Sim-Auto is its causally distilled streaming variant, generating successive chunks of camera and LiDAR observations as driving controls change, for long, interactive rollouts. Cosmos3-Nano-Sim-Transfer turns a physics simulator’s spatial controls — depth, segmentation, edges, and blur — plus a text prompt into photorealistic RGB video. Cosmos3-Nano-Sim-Bimanual predicts how a scene evolves in response to a two-arm robot’s actions, supporting interactive simulation for YAM and AgiBot embodiments.
NVIDIA also published the four-stage post-training recipe behind them — bidirectional adaptation, teacher forcing, consistency distillation, and self-forcing, where the student trains on its own generated history so compounding prediction errors are exposed during training rather than at deployment. Running the published setup requires Linux, an NVIDIA GPU, Docker, and the NVIDIA Container Toolkit; the documented vLLM-Omni configuration was tested on a single GB200.
Why it matters
World models are only useful to most teams if they can run them, adapt them to their own sensor rigs, and trust the license — and that is exactly what this release is aimed at. The checkpoint split is practical: generate training data offline with Transfer-Auto, distill it into Sim-Auto when you need a live loop, and use Sim-Bimanual to rehearse robot policies before risking real hardware. NVIDIA is candid about the open question that remains: whether these generated worlds faithfully support policy training and evaluation still has to be measured, not assumed. But as open weights with recipes included, they give researchers a shared starting point instead of another demo video.
Frequently asked questions
Can I use the Cosmos 3 Nano checkpoints commercially?
Yes, under the OpenMDW 1.1 terms, which permit commercial and non-commercial use subject to the license conditions.
What hardware do I need?
NVIDIA’s published setup calls for Linux, an NVIDIA GPU, Docker, and the NVIDIA Container Toolkit. The documented vLLM-Omni setup was tested on one GB200 — that is a tested configuration, not a stated minimum.
Are these finished self-driving or robot products?
No. NVIDIA presents them as post-trained examples of Cosmos 3 Nano — starting points for adaptation and research, with the reliability of generated worlds for policy training still to be measured.
Sources: NVIDIA via Hugging Face Blog, RuntimeWire