NVIDIA Releases Four Open-Weight Cosmos 3 Nano Checkpoints for Interactive Simulation
On November 5, 2026, NVIDIA added four post-trained Cosmos 3 Nano (16B) open-weight checkpoints under OpenMDW 1.1 for autonomous driving and robotics simulation.
NVIDIA added four open-weight checkpoints to the Cosmos 3 Nano family on November 5, 2026, targeting interactive world simulation: multi-sensor autonomous driving and bimanual robotics. The announcement came through a Hugging Face blog post; the weights are available for download under OpenMDW 1.1.
Four Roles, One Shared 16B Nano Format
All of them are presented as post-trained instances of Cosmos 3 Nano, featuring 16 billion trainable parameters:
- Cosmos3-Nano-Transfer-Auto — offline driving scene generation: up to 11 camera views (including fisheye), joint RGB–LiDAR generation, World Scenario Map controls, and text prompts for weather, lighting, and time of day.
- Cosmos3-Nano-Sim-Auto — distilled streaming variant: small chunks of synchronized observations for long, interactive rollouts.
- Cosmos3-Nano-Sim-Transfer — the physics simulator handles motion and contacts, while the model renders the RGB observation from spatial controls (depth, segmentation, edges, blur) + text.
- Cosmos3-Nano-Sim-Bimanual — predicting the next video frames from an image and robot actions (or camera movement); embodiment examples include YAM and AgiBot.
The common thread: connecting controls (scenario maps, actions, simulator signals) to generated observations, using chunk-based generation and an attention cache for sequential rollouts.
Licensing and Adaptation Pipeline
NVIDIA announced the OpenMDW 1.1 license, which states that it permits commercial and non-commercial use under its terms. The Transfer-Auto card does list openmdw1.1-license. The post details a four-stage recipe (bidirectional adaptation, teacher forcing, distillation, self-forcing) to adapt a checkpoint to a custom sensor configuration or embodiment—Transfer-Auto first, then distillation into Sim-Auto if streaming is needed.
The runtime remains demanding: Linux, NVIDIA GPU, Docker / NVIDIA Container Toolkit; a vLLM-Omni setup is cited as tested on a GB200, without an official minimum memory baseline.
Our take
Cosmos 3 Nano is not competing with consumer video generators: it is a world model building block for robotics and ADAS labs looking for synchronized multi-view and closed-loop workflows. Its practical value lies in the clear division of roles (offline data, streaming, rendering from physics simulators, bimanual actions) and open weights under OpenMDW 1.1. The real measure of adoption will play out far beyond the Hugging Face page: inference costs per rollout hour, stability over long trajectories, and actual gains when post-training on your own sensor fleet. For teams already on Isaac or in-house simulators, these four checkpoints offer a shared baseline for comparison—provided you read OpenMDW 1.1 before shipping a product.
Sources
- Hugging Face Blog — New NVIDIA Cosmos 3 Open Models for Interactive World Simulation (November 5, 2026) Primary source
- Hugging Face — nvidia/Cosmos3-Nano-Transfer-Auto Primary source
- OpenMDW 1.1 — License text Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
