NVIDIA has put four open-weight Cosmos 3 checkpoints on Hugging Face, and anyone with a suitable GPU can download them, including for commercial work. All four are post-trained from Cosmos 3 Nano and carry 16 billion trainable parameters. Two target driving and two target robot arms, according to NVIDIA’s post on the Hugging Face blog dated 30 October.
The license is OpenMDW 1.1, which NVIDIA says covers both commercial and research use, subject to its terms. The post does not summarize those terms. A company planning to ship a product on top of these weights will need legal to read the license itself.
Cosmos3-Nano-Transfer-Auto is the offline driving model. It produces synchronized video for up to 11 cameras, fisheyes included, and can generate LiDAR readings alongside the images. Per-camera controls fix the scene layout and how traffic moves, while a text prompt sets the look. A team can therefore replay one scenario in snow, at dusk, or in glare without recording any of it. LiDAR output needs extra map-derived range controls lined up with the camera ones.
The Sim-Auto checkpoint is that same model distilled for streaming. It generates camera and LiDAR in small chunks, a few denoising steps each, using its own earlier output plus fresh driving inputs. The checkpoint does not run a simulation by itself. Closed-loop testing needs a separate simulation runtime and a renderer that turns driving inputs into the control maps. Anyone with an unusual sensor layout is expected to adapt Transfer-Auto to it first and distill afterward.
The two robotics checkpoints take opposite approaches. Cosmos3-Nano-Sim-Transfer sits behind a physics simulator. The simulator keeps motion, contacts, and ground truth, and the model reads depth, segmentation, edge, or blur maps plus a prompt to paint color frames a policy can consume. Cosmos3-Nano-Sim-Bimanual has no simulator behind it. Given an image and a stream of actions from a policy or a human with joysticks, it predicts the next frames for two-armed YAM and AgiBot machines.
Bimanual comes with strict bookkeeping. The mixed-action example packs 59 control values, which training pads out to 64, and the schema spells out coordinate frames, rotations, and gripper calibration. Timing has to match too: every set of T transitions needs one more observation than that.
Running any of this means Linux, a GPU from NVIDIA, Docker, and NVIDIA’s container toolkit. NVIDIA tested its published setup on a single GB200 and says that figure is no minimum for memory. A university lab therefore cannot tell from the post whether its own card is enough.
The post also lays out the recipe for building the streaming models, in four stages. Training starts on full recorded clips, then predicts one chunk at a time from recorded history, then distills a faster student that needs four denoising steps per chunk. The last stage trains the student on its own generated history, because a slightly wrong object in one frame becomes context for the next and errors snowball. NVIDIA recommends handing the matching skill file and a dataset location to a coding agent, and warns that a short successful run proves the pipeline executes, not that training has converged.
NVIDIA’s own wording is candid about the limits. It says the remaining job is to find out how far the models can be trusted for training and judging robot policies. The post includes no benchmark scores and no comparison with other simulators, so fidelity is still the open question. Until someone publishes a test of a policy trained on generated frames and then run on a real car or arm, these checkpoints are a free starting point and carry no proof that the skills transfer.
Reported by the NVIDIA post on the Hugging Face blog on 30 October 2026.
