Nvidia has expanded its Cosmos 3 world-model platform with four open-weight checkpoints covering autonomous driving and robotic manipulation simulation. Each 16-billion-parameter model is available on Hugging Face under a license that permits commercial use — no access approval required.
Nvidia on Nov. 5, 2026 released four new open-weight model checkpoints under its Cosmos 3 world-model development platform, extending the family’s reach into two of the most data-hungry domains in physical AI: autonomous vehicle simulation and robotic manipulation. All four models are post-trained specializations of the same Cosmos 3 Nano base, each carrying 16 billion trainable parameters and released under the OpenMDW1.1 license, which allows both commercial and non-commercial use without field-of-use restrictions.
The four checkpoints are Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer and Cosmos3-Nano-Sim-Bimanual. Rather than general-purpose world generators, each one solves a specific simulation problem. Transfer-Auto produces synchronized video across up to 11 camera views — including fisheye lenses — alongside joint RGB and LiDAR data, making it useful for offline synthetic driving dataset creation with controllable weather, lighting and time of day. Sim-Auto is the streaming variant of Transfer-Auto, generating successive short video-and-LiDAR chunks in a few denoising steps so a driving policy can interact with the simulation in something close to real time.
The robotics-focused pair works differently. Sim-Transfer takes spatial control signals from a conventional physics simulator — depth maps, segmentation masks, edge maps or blurred scene previews — and renders photorealistic RGB frames that a robot policy can act on. The simulator retains responsibility for ground-truth physics; Cosmos handles appearance. Sim-Bimanual skips the physics simulator entirely and predicts the next visual observation directly from a two-arm robot’s actions, supporting interactive bimanual simulation for YAM and AgiBot hardware under either policy or joystick control.
What Makes Cosmos 3 Different From Earlier Releases
The architectural shift in Cosmos 3 is its Mixture-of-Transformers (MoT) backbone, which connects separate text-processing and visual-generation networks through shared attention. Earlier Cosmos releases required developers to work with distinct models for world generation, controlled generation, scene understanding and policy generation. The new unified base consolidates those capabilities, and the four checkpoints announced on Nov. 5 represent a further specialization layer on top of it — ready-made starting points for specific tasks rather than blank slates.
The post-training pipeline follows a four-stage recipe: bidirectional adaptation to learn domain and control relationships, teacher forcing to predict chunks from recorded observations, consistency distillation to reduce the number of denoising steps needed, and self-forcing to handle the gap between generated and real history during streaming rollouts.
How This Release Fits the Competitive Landscape
Several well-resourced teams are working on world models for physical AI, but the competitive map is fragmented. Google DeepMind’s Genie 3 targets interactive text-to-world generation for training curricula; Meta’s V-JEPA 2 is a self-supervised video model aimed at zero-shot robotic planning; and Wayve’s GAIA-2 focuses narrowly on controllable multi-camera driving video for safety-critical scenarios. Critically, Genie 3 and World Labs’ Marble are closed or subscription-gated, while Nvidia Cosmos ships as open weights on Hugging Face and NGC. No major competitor currently offers ungated coverage of both autonomous vehicle and bimanual manipulation simulation under a single model family — that simultaneous breadth is the key differentiator here. Google still holds the stronger position in general-purpose world-model research overall, but Nvidia’s tight integration with Isaac Sim and the GR00T robot foundation model gives Cosmos 3 a commercial deployment path that academic releases lack.
Why Students and Early-Career Researchers Should Pay Attention
Physical AI research has historically been gated by two expensive bottlenecks: real-world data collection and physical hardware. Cosmos 3 addresses both. A student working on robot manipulation can use Sim-Bimanual to generate training demonstrations without a physical robot arm or teleoperation rig. A student in an autonomous vehicles course can use Transfer-Auto to synthesize multi-camera driving datasets with rare weather conditions that no public benchmark covers. Because the weights are not gated and the license places no obligations on model outputs, student projects can include commercial applications or startup pilots from day one.
The practical ceiling is GPU access. Inference requires Linux, an Nvidia GPU, Docker and the Nvidia Container Toolkit. The technical report’s reference setup used a GB200, though Nvidia notes that is a test configuration rather than a minimum specification. Students without lab hardware will likely need a cloud GPU provider or access to a university HPC cluster.
On the job market, familiarity with the Cosmos ecosystem is increasingly a differentiator. Companies such as FieldAI and Skild AI are actively building generalized robot intelligence on top of Nvidia Cosmos world models, and the broader Cosmos Coalition includes hiring employers like ABB Robotics, FANUC, Figure and Universal Robots. Getting hands-on time with these checkpoints now — running inference, fine-tuning on a small custom dataset, contributing to open benchmarks — puts students ahead of a credential curve that hasn’t fully formed yet.
Model weights, technical report and demo notebooks are available via the Nvidia announcement on Hugging Face.
Source: Hugging Face