NVIDIA just dropped Cosmos 3 — a single model that thinks, predicts, and acts in the physical world. No more stitching together separate AI components for vision, reasoning, and control. Cosmos 3 bundles it all into one unified system.

Built on a novel mixture-of-transformers architecture, Cosmos 3 pairs a reasoning tower with a generation tower. The reasoner interprets multimodal inputs — video, images, text, and actions — while the generator produces future scenarios, video sequences, and action trajectories. This combo handles everything from understanding a scene to predicting and generating robot movements.

Before Cosmos 3, developers juggled multiple models for world generation, policy learning, and physical reasoning. Now, one model streamlines workflows, cutting down complexity and inference overhead. Whether simulating autonomous driving scenarios or robotic pick-and-place tasks, Cosmos 3 delivers realistic, physically plausible outputs in a single forward pass.

The model comes in two sizes. Cosmos 3 Nano packs 16 billion parameters optimized for real-time robotics inference on workstation GPUs like the RTX PRO 6000. Cosmos 3 Super is a larger, 64 billion parameter beast aimed at datacenter-scale synthetic data generation and research, running on NVIDIA’s Hopper and Blackwell GPUs.

Cosmos 3 supports diverse input-output pairs—text, image, video, and action—allowing it to serve as a vision-language model, video world model, or action policy model. This flexibility means developers can generate rare, long-tail scenarios for testing and training robots or autonomous vehicles without costly real-world data collection.

Open-source synthetic datasets released alongside Cosmos 3 cover robotics, spatial reasoning, human motion, autonomous driving, and warehouse safety. These datasets help developers fine-tune the model for specific domains or embodiments, accelerating adoption across industries.

Deployment is simplified with NVIDIA’s NIM microservices, delivering optimized inference runtimes. These microservices support quantized model checkpoints, including a 4-bit floating point format, doubling inference speed without sacrificing accuracy. This makes Cosmos 3 practical for production environments, not just research labs.

Cosmos 3 leads on multiple physical AI benchmarks. It tops vision reasoning leaderboards like VANTAGE-Bench and the Traffic Anomaly Reasoning challenge. It also outperforms competitors on video generation and policy benchmarks such as Artificial Analysis, PAI-Bench, Physics-IQ, and RoboLab.

The model’s reasoning capabilities let it simulate complex motions, spatial relationships, and causality, enabling AI agents to “think before they act.” This is crucial for robots navigating unpredictable environments or autonomous vehicles facing rare edge cases.

NVIDIA also formed the Cosmos Coalition, a global alliance of AI labs and robotics companies collaborating on open world models. This coalition fosters shared innovation, interoperability, and faster progress in physical AI development.

Cosmos 3 signals a shift from scripted robot routines to embodied autonomy. It bridges the gap between simulation and real-world deployment by providing a generalist physical AI foundation. The goal: robots and AI agents that understand and predict their environments, reducing failures and enabling safer, smarter automation.

Developers can access Cosmos 3 models, training scripts, and datasets on Hugging Face and GitHub. The platform’s open licensing under OpenMDW 1.1 encourages experimentation, customization, and redistribution across physical AI workflows.

This release isn’t just another AI model. It’s a bold bet that the next AI frontier lies in embodied, physical intelligence. And NVIDIA is staking a claim at the center of that future.