Nvidia released four open-weight checkpoints for its Cosmos 3 world-model family on November 5, 2026, giving developers new tools to simulate driving and robotics in software before anything touches the real world. The models generate interactive, physically plausible scenes — synchronized camera video and LiDAR, robot manipulation, and photorealistic video from simulator controls — under a license that permits commercial use.
What a world model is
A world model is an AI system that predicts how a physical scene looks and changes as actions unfold. Instead of filming every possible driving scenario or running every robot trial in a warehouse, developers can ask a world model to generate the scene: a rainy intersection at dusk, a forklift crossing a lane, a robot arm reaching for a part. Cosmos is Nvidia's platform for building such models, pairing the models themselves with developer tools and the GPU infrastructure to run them.
The four new checkpoints
All four checkpoints are post-trained versions of Cosmos 3 Nano, each with 16 billion trainable parameters, released under the OpenMDW1.1 license, which permits commercial and non-commercial use under its terms.
Cosmos3-Nano-Transfer-Auto is the offline workhorse. A bidirectional model for synthetic driving-data generation, it produces synchronized video across up to 11 camera views — including fisheye lenses — together with matching LiDAR point clouds. Developers can set per-camera World Scenario Map parameters to define road layout and traffic behavior, and use text prompts to steer weather, lighting, and time of day.
Cosmos3-Nano-Sim-Auto is the streaming version of Transfer-Auto. Causally distilled for real-time use, it generates successive small chunks of synchronized camera video and LiDAR in a few denoising steps, enabling interactive driving simulation rather than batch generation.
Cosmos3-Nano-Sim-Transfer takes a physics simulator's controls — depth, segmentation, edges, blur — plus a text prompt, and turns them into photorealistic video observations. It bridges the gap between a simulator's abstract state and the rich visual input a perception system needs to learn from.
Cosmos3-Nano-Sim-Bimanual predicts video from robot or camera actions, aimed at manipulation tasks where a robot must foresee the visual consequences of its own movements. It supports interactive two-arm simulation for YAM and AgiBot robot embodiments, controlled by a learned policy or human-operated joysticks.
Why open weights matter
The headline is not the parameter count — 16 billion is modest by 2026 standards — but the open-weight, commercial-use release. The ability to synthesize synchronized multi-camera video and LiDAR together has until now lived mostly inside closed, proprietary systems at large autonomous-vehicle companies. Releasing it openly lowers the cost of building perception and planning systems for smaller teams, and it makes simulation a shared infrastructure rather than a competitive moat held by a few giants. The models run on Nvidia's Hopper and Blackwell GPUs, and Nvidia's technical blog frames the platform as serving physical AI across autonomous vehicles, robotics, and smart infrastructure.
Why it matters for founders
For founders in autonomy, robotics, and embodied AI, the bottleneck has shifted. Data collection — fleets of sensor-equipped vehicles, warehouse trials, edge-case hunting — is expensive and slow. High-quality synthetic data generated by world models compresses that cycle: rare and dangerous scenarios can be generated on demand, policies can be trained against thousands of variations, and iteration can happen in simulation before a single real-world mile. The caution is that generated data is only as good as its fidelity to the real world; teams still need to validate against real sensor data, and Nvidia's own materials offer no independent comparison of generated LiDAR against physical sensors. Founders who build rigorous real-versus-simulated validation into their pipelines will capture the speed advantage without inheriting the blind spots.
Angeled.com connects founders with investors.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



