NVIDIA's Cosmos 3 Nano Opens Four World Models for Driving and Robots — All 16B, All Open Weight

NVIDIA Opens Four Cosmos 3 Nano World Models

On 5 November 2026, NVIDIA released four open-weight checkpoints from its Cosmos 3 Nano family, each with 16 billion trainable parameters, under the OpenMDW 1.1 licence, which permits commercial and non-commercial use under its terms. Two generate driving worlds, with synchronised video across up to 11 camera views and matching LiDAR; the other two turn simulator controls and robot actions into video. A world foundation model is an AI model trained to simulate how the physical world looks and behaves, so robots and driving systems can practise inside a generated world before they act in the real one.

NVIDIA is unusually honest about what is still unproven: its own announcement says the hard work ahead is measuring where these models faithfully support policy learning and evaluation. No one outside NVIDIA has independently measured these four checkpoints yet.

The Four Checkpoints, and What Each One Does

The four checkpoints split into two families. The driving pair handles cars; the robotics pair handles simulation and manipulation. All four are post-trained versions of the same Cosmos 3 Nano base model, which is why they share the 16 billion trainable parameter figure and the same licence.

Checkpoint	Job	What it generates
Cosmos3-Nano-Transfer-Auto	Offline synthetic driving data	Synchronised video across up to 11 camera views, including fisheyes, plus joint RGB and LiDAR generation
Cosmos3-Nano-Sim-Auto	Streaming driving simulation	Successive small chunks of synchronised camera video and LiDAR in a few denoising steps, for long interactive rollouts
Cosmos3-Nano-Sim-Transfer	Simulator to video	Photorealistic RGB observations from a physics simulator's spatial controls and a text prompt
Cosmos3-Nano-Sim-Bimanual	Robot action to video	Video predicting how a scene changes when a two-armed robot acts, for YAM and AgiBot embodiments

Transfer-Auto is the offline workhorse. It is a bidirectional model, which means it looks at a whole recorded clip at once, and it generates matched camera footage and LiDAR point clouds under controls for scene structure, traffic motion, weather, lighting and time of day. Developers can also give it an initial image to anchor how the scene looks, and aligned HD-map range controls guide the LiDAR output. Sim-Auto is its streaming counterpart: a causally distilled version that produces the next small chunk of video and LiDAR from generated history and updated driving controls, so a simulation can keep running and reacting instead of rendering one finished clip. Closed-loop use needs a simulation runtime and a renderer to update the world controls from driving inputs.

Illustration created for INORYVA

Sim-Transfer works the other way round from a normal video model. Your physics simulator computes the motion and the physical state, and Sim-Transfer only renders what the robot's camera would see, from spatial controls for depth, segmentation, edges and blur plus a text prompt. The physics stays with the simulator, which is the point: the video is a view of a simulated world, not a guess at one. Sim-Bimanual predicts the next observation directly from a two-armed robot's actions, driven by a policy or by a person with joysticks.

16 Billion Parameters — and the 15 Billion Listing

Readers who check the Hugging Face model pages may notice a number that does not match: at least one listing displays a rounded figure near 15 billion. NVIDIA's announcement says all four checkpoints have 16 billion trainable parameters, and the base Cosmos 3 Nano is a 16 billion class model built from an 8 billion parameter reasoner and an 8 billion parameter generator. The listing figure is a display rounding, not a second version of the model. Cosmos 3 is built on a Mixture-of-Transformers design, pairing an autoregressive reasoner with a diffusion generator, so the model reasons before it generates.

The four checkpoints were trained with a four-stage recipe that NVIDIA also published, ending in self-forcing, where the student trains on its own generated history so small errors do not compound into broken rollouts. The post-training recipes for all four variants are public.

What Open Weights Change Here

Open weights do not mean the model is free to run, and they do not mean it is good. They mean the weights are downloadable and the licence permits commercial use, so a driving team can adapt a checkpoint to its own camera layout and LiDAR configuration instead of starting from scratch or buying access to a closed system. Transfer-Auto is explicitly positioned as the starting point for that adaptation: post-train it on your own driving data, then distil it into Sim-Auto for fast streaming.

What You Need to Run Them

The published requirements are Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit. NVIDIA's published vLLM-Omni setup was tested on one GB200, and the company states plainly that this is not a minimum-memory specification, so do not read it as a hardware requirement. Inference runs through open skills and notebooks linked from the announcement. There is no price to quote, because open weights carry no per-use fee. The cost is hardware, engineering time and the data needed to adapt the model to your own domain.

Illustration created for INORYVA

Cosmos Is Not Genie 4

Because both are called world models, it is easy to confuse this release with Google DeepMind Genie 4, announced on 27 October 2026. Genie 4 is a closed research preview for trusted testers that generates an interactive world you step into, with native audio, music and dialogue, and it is coming to Google AI Ultra subscribers and early Gemini API partners. You cannot download its weights. Cosmos 3 Nano is the industrial branch: open weights for generating synthetic training data and simulating futures for robots and autonomous vehicles, with no consumer experience attached. One is a world you explore; the other is a world you train in.

The open-weight peer worth comparing is Mistral Large 4 open weights, a language model release rather than a world model, but the same trade applies: downloadable weights, a permissive licence, and the responsibility to evaluate the model yourself before you rely on it.

Who Should Try It, and Who Should Wait

Autonomous-driving teams with multi-camera rigs and LiDAR should look at Transfer-Auto first, because it directly addresses their hardest data problem: matched video and geometry across many views, under varied conditions. Robotics teams working with two-armed manipulation should look at Sim-Bimanual, and teams with an existing physics simulator should look at Sim-Transfer. Researchers in world-model evaluation have a new open baseline to test against.

Everyone else should wait. If you do not run GPU infrastructure or work with driving or robotics data, these checkpoints are not for you. There is no demo to try in a browser and no hosted version announced.

Illustration created for INORYVA

For the wider picture, see our flagship AI model comparison.

Key Takeaways

• NVIDIA released four open-weight Cosmos 3 Nano checkpoints on 5 November 2026, each with 16 billion trainable parameters under the OpenMDW 1.1 licence.
• Transfer-Auto generates offline multi-view driving data with joint RGB and LiDAR; Sim-Auto streams it; Sim-Transfer renders simulator controls into video; Sim-Bimanual predicts video from two-armed robot actions.
• The 15 billion figure on one model listing is a display rounding; NVIDIA's announcement states 16 billion trainable parameters.
• No independent benchmarks exist yet, and NVIDIA says measuring where the models support policy learning is unfinished work — test on your own data before relying on them.
• Running them needs Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit; the GB200 test setup is not a minimum-memory specification.
• Cosmos is the industrial, open-weight branch of world models; Genie 4 is the closed, consumer-facing branch. They solve different problems.

Conclusion

NVIDIA's four Cosmos 3 Nano checkpoints give driving and robotics teams something they did not have a week ago: downloadable, commercially usable world models for multi-view driving data, streaming simulation and robot action prediction. The release is strongest where it is most honest, with public recipes, a clear licence and an open admission that evaluation is unfinished. The teams who benefit most will be the ones who adapt the checkpoints to their own sensors and domains, and measure the results themselves.

Sources

• NVIDIA, "New NVIDIA Cosmos 3 Open Models for Interactive World Simulation", Hugging Face blog, published 5 Nov 2026 — https://huggingface.co/blog/nvidia/cosmos-interactive-world-simulation
• NVIDIA Cosmos 3 collection, Hugging Face — https://huggingface.co/nvidia
• NVIDIA, "Welcome NVIDIA Cosmos 3: The First Open Omni-model for Physical AI Reasoning and Action", Hugging Face blog — https://huggingface.co/blog/nvidia/cosmos-3-for-physical-ai