What happened
NVIDIA presents Cosmos3-Nano as one model in Cosmos3, a collection of omnimodal world models for Physical AI applications and research. Its model card says the family can generate dynamic video, images, audio and action commands from combinations of text, image, video and action-trajectory inputs. The stated use cases include world understanding and generation, simulation, future prediction, action reasoning and embodied-policy learning. NVIDIA identifies robotics, autonomous vehicles and smart-space settings, including industrial and factory-scale applications, as target environments.
Cosmos3-Nano has 16 billion trainable parameters. The card describes a Mixture-of-Transformers architecture with two transformer towers: an autoregressive transformer for discrete-token generation and a diffusion transformer for continuous multimodal generation. Text uses next-token decoding; non-text outputs such as images, video, audio and actions are synthesized through iterative denoising. NVIDIA says the model is based on the Cosmos Framework.
The model card lists text, image, video, audio and action trajectories among generator inputs. Input videos can be 256p, 480p or 720p, with no more than five frames; text input is limited to 4,096 tokens. Audio input is limited to half a second. Action trajectories are sequences of per-frame state or control values, with lengths from 16 to 400 frames. Supported embodiments include general camera motion and autonomous-vehicle motion (both 9D), egocentric motion (57D), and several robot configurations, including single- and dual-arm Franka Panda, Agibot, UR, Google robot, WidowX 250 and UMI.
Outputs can include images, video, audio, action and text. Generated videos range from 5 to 400 frames, with 189 frames as the default; audio is stereo AAC muxed into MP4. Action output is limited to the compatible embodiments listed in the card. A separate reasoner accepts text, text plus image, or text plus video and returns text; its context window is listed as up to 256K tokens. NVIDIA lists PyTorch, vLLM-Omni, Hugging Face Diffusers and SGLang integrations, and NVIDIA Ampere, Blackwell and Hopper GPUs on Linux. The card says only BF16 precision is tested. It lists an OpenMDW 1.1 license and says the model is ready for commercial and non-commercial use, subject to that license. It gives May 31, 2026 as the release date on Hugging Face and GitHub.
What the archive adds
NVIDIA’s October 21, 2026 Cosmos platform expansion added training tools, simulation checkpoints and Cosmos Data Factory blueprints, as New Runtime’s earlier coverage explains. A later Cosmos Framework note added implementation details on synchronized camera-and-LiDAR recipes, model families and inference. Cosmos3-Nano is a new model release within that established platform thread, not merely a restatement of those platform announcements.
Trend signal
This does not yet fit a pattern New Runtime tracks. The model’s unified handling of simulated worlds, media and actions is relevant to Physical AI, but the supplied pattern set does not establish a sufficiently close thesis to claim that this release supports one.
Why now
NVIDIA’s model card dates the Hugging Face and GitHub release to May 31, 2026. That makes the published weights and implementation details the immediate development milestone; the card does not explain why NVIDIA chose that timing.
Why it matters
For teams building robots or autonomous systems, a world model that accepts proposed controls and can produce action outputs creates a concrete component to evaluate alongside perception and simulation tools. The listed embodiment dimensions matter: an action sequence for one robot configuration is not automatically interchangeable with another, and the card’s supported formats define the boundary of what builders can test directly.
The release also makes the deployment trade-off visible. The model is a 16B-parameter component designed for NVIDIA GPU-accelerated systems, with Linux support and BF16 as the only tested precision. Teams should therefore account for hardware and precision constraints before treating it as a lightweight general-purpose model. NVIDIA explicitly says use-case-specific testing is needed before integration and deployment. The model card establishes capability and access claims, not evidence that a robot trained with it will be safe or reliable in a particular environment.
Limits and uncertainty
NVIDIA’s model card lists input and output formats, resolution and sequence bounds, supported embodiments, runtime integrations and tested hardware. Those specifications do not establish performance on a particular robot, transfer across embodiments, or reliability in real-world operation. The card calls for use-case-specific testing and validation before deployment. It does not, in the material described here, provide an independently checked assessment of operational safety or a guarantee that generated actions meet a team’s control requirements. The license is named, but builders should review its terms for their intended use.
Skill to build now
Build a small compatibility test around one supported embodiment: prepare a short action trajectory with the required per-frame dimensions, pair it with the relevant visual input, and check whether the generated output matches the representation your simulator or evaluation pipeline expects. Record the GPU, BF16 setting and input limits alongside results so the team can distinguish model behavior from integration constraints.
Reader takeaway
Cosmos3-Nano is a released, 16B-parameter model for multimodal Physical AI work, with explicit support for action trajectories and outputs. Its value to a builder is a testable interface across perception, simulation and action—not proof of a production-ready robot policy. Start with the exact embodiment and hardware constraints in NVIDIA’s model card.