NVIDIA posts four open Cosmos 3 Nano world models

NVIDIA announced four open-weight checkpoints for its Cosmos 3 world-model family in a Hugging Face blog post dated November 5, 2026. The four are Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer and Cosmos3-Nano-Sim-Bimanual. NVIDIA describes each as a post-trained example of Cosmos 3 Nano with "16B trainable parameters," released under the OpenMDW-1.1 license.

The checkpoints are not video generators aimed at creators. Each one produces the visual observations a robot or an autonomous vehicle would receive, conditioned on the actions and scene controls it is handed. They are tools for building and testing Physical AI systems, and NVIDIA's own documentation is firm about where their accuracy ends.

What the four checkpoints do

NVIDIA split the four across two jobs: driving and robotics. Cosmos3-Nano-Transfer-Auto generates synchronized multi-camera driving video, up to 11 standard and fisheye views, with optional matching LiDAR. It is a bidirectional model, which means it denoises a whole multi-view clip at once, so NVIDIA positions it for offline synthetic data. Its inputs are per-camera World Scenario Map controls, which encode road layout, lane markings, traffic signs and object boxes, plus a text prompt for appearance. The model card says it produces up to 300 frames per view at 480p or 720p.

Cosmos3-Nano-Sim-Auto is the streaming version of that work. NVIDIA distilled Transfer-Auto into a few-step model that generates short chunks of camera video and LiDAR in a rolling loop, so a driving policy can see the next observation as it acts. Getting a closed loop, according to the blog, also needs a separate simulation runtime and a World Scenario Map renderer to turn driving inputs back into controls.

Cosmos3-Nano-Sim-Transfer runs the transfer in the other direction. A physics simulator such as NVIDIA Isaac Lab keeps the state and the physics, then hands over depth maps, segmentation masks, edge maps or blurred frames, and the model renders photorealistic RGB video that follows that geometry. The model card lists up to 600 frames at 480p and a design target of eight parallel environments on one GB200.

Cosmos3-Nano-Sim-Bimanual predicts the next frames from a two-arm robot's actions. Give it a starting image and a stream of robot and camera actions, and it generates the visual consequences in small chunks, in open-loop or closed-loop form. It supports the YAM and AgiBot robot embodiments, driven by a policy model or a human joystick.

What a world model is here

A world model predicts how a scene changes, rather than how a clip should look. The four checkpoints inherit that behavior from Cosmos 3, whose Mixture-of-Transformers backbone runs a text network and a visual network side by side. In these checkpoints the text prompt and the controls steer the generator, and a key-value cache carries earlier chunks forward so the next one stays consistent with the last.

The practical result is interaction. Feed in an action, get back an observation, feed the observation into the next action. NVIDIA argues this is what robotics and autonomous-driving work needs and what pre-recorded footage cannot give, because a policy can respond to the consequences of its own decisions. The blog is candid that the value is not yet proven. The work ahead, it says, is "to measure where the models faithfully support policy learning and evaluation, understand their limits, and improve where they fall short."

The numbers, and the GPUs behind them

The performance figures belong to NVIDIA and to named hardware.

The headline number is for Sim-Bimanual. NVIDIA's technical report, also on arXiv, reports 48.1 frames per second with median and 95th-percentile chunk times of 166 and 169 milliseconds, at 480p on a single GB300 GPU. That run used four denoising steps, a 64-latent-frame history window, and chunks of two latent frames decoded into eight RGB frames at 640×352. NVIDIA notes that at a 30 FPS playback rate an eight-frame chunk is 267 milliseconds of physical time, so generation runs ahead of real time on that workload.

The model card gives a second, different picture. Its own table of 480p inference throughput, measured with vLLM-Omni on one GPU, lists GB200 at 54.1 frames per second, H100 at 32.6, H200 at 32.4, RTX PRO 6000 at 20.6 and H20 at 10.7. Sim-Transfer's card lists lower figures on the same hardware set, with GB200 at 24.5 frames per second and RTX PRO 6000 at 8.3.

For driving, the developer blog says Transfer-Auto supports 30 fps input and output with up to 400 frames at 480p or 720p, while the Transfer-Auto model card caps output at 300 frames per view. NVIDIA's reference run produced seven 720p videos of 297 frames each, at 30 FPS, with 99 LiDAR sweeps.

These are NVIDIA's own comparisons. The project page says Sim-Bimanual posts higher video quality and reference-frame PSNR than DreamDojo 2B and 14B on two AgiBot benchmarks, without publishing the values, and that Sim-Transfer shows 59.2% higher edge F1 and 26.5% lower depth error than its Cosmos 3 reference. Automatica Press noted that the DreamDojo comparison gives no metric values and that the release carries no independent tests.

What OpenMDW-1.1 permits

The license is the OpenMDW License Agreement, version 1.1. Its grant is broad: permission "to deal in the Model Materials without restriction, including under all copyright, patent, database, and trade secret rights." It adds that it "does not impose any restrictions or obligations with respect to any use, modification, or sharing of any outputs generated by using the Model Materials."

Commercial use is allowed. NVIDIA's blog says the models "use OpenMDW1.1, permitting commercial and non-commercial use under its terms," and each model card states the model "is ready for commercial or non-commercial use."

Two conditions are worth knowing. If you distribute any part of the model, you must keep a copy of the agreement and the notices of origin. The license also ends if you sue anyone claiming the model materials infringe a patent or copyright, unless you are answering a suit first brought against you. The whole thing ships "as is," with no warranty, and it puts responsibility for clearing rights in inputs and outputs on you.

What the checkpoints cannot do

NVIDIA's Cosmos 3 Nano card is plain that the family is not a physics engine. Because the model "lacks an explicit physics simulator," the card states, "3D geometry, 4D space-time evolution, object permanence, contact dynamics, and physical laws are only approximated," which produces artifacts such as "disappearing or morphing objects, unrealistic collisions, and physically implausible motions." Its summary line is blunt: "Cosmos3 outputs should not be treated as physically accurate simulation, reliable ground-truth reasoning, or safety-certified decision making."

For robotics and driving, NVIDIA says these applications "require independent validation, system-level safety analysis, and appropriate constraints," and that the model "is not designed or tested for uses where failure could cause injury, death, or catastrophic damage."

The technical report documents failure modes for Sim-Transfer across all four control types: temporal degradation such as late-rollout ghosting, object hallucination, appearance drift, geometric deformation, and errors in following the supplied controls. The models also need serious hardware and a Linux, Docker and NVIDIA GPU stack. NVIDIA's deployment guide lists 58 GiB of GPU memory for Cosmos 3 Nano generation in BF16 and 44 GiB with FP8 and offloading.

What it means for creators

For a video creator reading this on the way to a client job, the four checkpoints are simulation and training-data tools, not a shot generator. There is no text box that returns a usable narrative clip, and none of these models is built to illustrate a script.

Where they could realistically fit a small studio: a team with robot or automotive clients could generate multi-view driving footage with matched LiDAR for previz or for a synthetic-data pipeline. A team running a physics simulator could use Sim-Transfer to change the look of a simulated scene while keeping its motion, which the developer blog describes as altering lighting, textures and backgrounds while preserving geometry. A studio that previzes shots with two-arm robots could use Sim-Bimanual to predict what a robot arm would do next from its actions.

All of that needs an NVIDIA datacenter GPU, a Linux host and Docker, and most creators would rent that compute rather than own it. The models generate camera, LiDAR or robot-action video rather than an editable timeline. Any output shown to a client should be labeled for what it is, because NVIDIA's own card says it is not physically accurate.

NVIDIA frames the release as a shared starting point, with open weights and post-training recipes that it says are for "adapting models, comparing methods, and contributing improvements." The earlier SLOP TV story on Cosmos3-H-Dreams, the surgical simulator NVIDIA shipped on October 20, 2026, covers a different and narrower Cosmos 3 product; the four checkpoints here are the general simulation recipe.

The four checkpoints are on Hugging Face under the OpenMDW-1.1 license, with code at github.com/nvidia/cosmos and github.com/nvidia/cosmos-framework.