NVIDIA Cosmos 3: Open Source Physical AI Explained (2026)

Building a robot that can actually navigate the real world used to require three separate AI pipelines: one to understand the scene, one to simulate what might happen next, and one to generate the right action. Coordinating those systems meant latency spikes, engineering overhead, and models that broke down whenever conditions got unpredictable. On June 1, 2026, NVIDIA replaced that stack with a single open model. Cosmos 3 combines physical reasoning, world generation, and action prediction into one architecture — and released it free to anyone with a Hugging Face account. For robotics developers, autonomous vehicle teams, and researchers on tight budgets, that is a significant shift.

What Is NVIDIA Cosmos 3 and Why Does It Matter?

Cosmos 3 is the third generation of NVIDIA's physical AI world foundation model series. According to the official NVIDIA newsroom announcement, it is "the world's first fully open omnimodel" that can natively understand and generate text, images, video, ambient sound, and physical action sequences — all within a single system.

Previous physical AI systems typically required separate models for perception (identifying what is in a scene), prediction (simulating what will happen next), and control (generating the action to take). Cosmos 3 unifies all three. The model's architecture pairs a reasoning component that analyzes spatial and temporal context with a generation component that produces future observations and action trajectories based on that analysis.

The practical result, according to NVIDIA, is a reduction in physical AI training and evaluation cycles from months to days. For a field where collecting real-world robot training data is expensive, slow, and sometimes dangerous, that compression matters enormously. Cosmos 3 can generate rare or hazardous scenarios — robot collisions, unusual road events, edge-case warehouse situations — synthetically, without putting hardware or people at risk.

Jensen Huang, NVIDIA's founder and CEO, described the significance at the launch: "The Cosmos 3 family of open, frontier omnimodels gives developers a generational leap in ability to build robots, autonomous vehicles and vision AI that perceive, reason, plan and act in the physical world."

What Are the Two Cosmos 3 Variants — Nano and Super?

NVIDIA released Cosmos 3 in two sizes, each targeting a different deployment context:

The Nano variant is the more accessible of the two. At 16 billion parameters, it is designed to run on workstation-class professional hardware, meaning researchers and small teams do not necessarily need a full cloud datacenter to experiment. The Super variant at 64 billion parameters delivers maximum accuracy and is optimized for synthetic data pipelines and high-fidelity physical simulation — the kind of workloads that typically run in large-scale training environments.

Both models share the same Mixture-of-Transformers (MoT) architecture and support the same input and output modalities: text, images, video, ambient sound, and action sequences. The difference is in capacity and the hardware tier required to run them at full performance.

How Does the Mixture-of-Transformers Architecture Work?

The core architectural innovation in Cosmos 3 is what NVIDIA calls Mixture-of-Transformers (MoT). This is a dual-tower design that routes reasoning and generation through specialized components, rather than trying to handle everything with a single undifferentiated transformer stack.

The Reasoner Tower acts as the model's perception and analysis layer. It is a vision-language model that processes incoming images, video frames, and text to build an understanding of the scene: what objects are present, how they are moving, what physical constraints apply, and what is likely to happen next. Think of it as the model's "thinking" phase — it interprets physical context before any generation begins.

The Generator Tower then takes that physical context and produces outputs. Using diffusion-based techniques, it generates future video frames and action sequences — effectively simulating what the world will look like after a given action, or producing the optimal action for a given situation. The generator tower does not start from scratch; it conditions its outputs on the reasoning tower's analysis, which is what makes Cosmos 3's predictions physically consistent rather than visually plausible but physically wrong.

As NVIDIA's official blog explains, the model was trained on one of the largest multimodal physical AI datasets ever assembled, including billions of samples across text, image, video, sound, and action trajectories. That breadth of training is what allows Cosmos 3 to generalize across different robot form factors, vehicle types, and physical environments without requiring massive amounts of task-specific fine-tuning data.

How Can Developers Access Cosmos 3 for Free Right Now?

NVIDIA has made Cosmos 3 available through multiple free access paths simultaneously:

• Hugging Face — Both Cosmos 3 Nano and Cosmos 3 Super are available in the nvidia/cosmos3 collection on Hugging Face. You can browse model cards, download weights, and explore documentation directly from the hub.
• GitHub — Training scripts, fine-tuning code, synthetic data generation pipelines, and inference examples are all available at the nvidia/cosmos GitHub repository. This is the starting point for anyone who wants to adapt Cosmos 3 to a specific robot or vehicle platform.
• build.nvidia.com — NVIDIA's developer playground lets you test Cosmos 3 in your browser without downloading or installing anything. This is useful for initial evaluation before committing to a local or cloud deployment.
• NVIDIA NIM microservices — For teams that want managed inference at scale, Cosmos 3 is also available as packaged NIM microservices, which simplify deployment in cloud and enterprise environments.

The license governing all these access paths is OpenMDW 1.1, a model license developed by the Linux Foundation specifically for physical AI applications. It allows free use for both research and commercial purposes, with attribution requirements. This is meaningfully more permissive than many competing physical AI model releases, which often restrict commercial use or require separate enterprise licensing agreements.

If you are comparing open source AI options for your project, our open source AI model directory tracks the latest releases across robotics, language, and multimodal categories.

What Is the NVIDIA Cosmos Coalition and Who Is Involved?

Alongside the Cosmos 3 model release, NVIDIA announced the formation of the NVIDIA Cosmos Coalition — a collaborative group of AI labs and robotics companies working together to advance open world models for physical AI.

The founding coalition members announced at launch include:

• Agile Robots — Advanced humanoid and collaborative robot platforms
• Black Forest Labs — Frontier generative AI research
• Generalist — General-purpose robot learning systems
• LTX — Video generation and simulation
• Runway — AI-native video and creative tools
• Skild AI — Robot learning and generalist robot policies

The coalition structure matters because it signals that Cosmos 3 is not just a model drop — it is intended to become a shared foundation that multiple organizations build on, contribute fine-tuned variants to, and extend for specialized domains. This mirrors the ecosystem dynamics that made models like Llama 3 and Mistral successful: a strong base model with an open license creates a flywheel of community improvement.

For developers choosing between building on Cosmos 3 and proprietary physical AI platforms, the coalition membership is a meaningful signal about the ecosystem you are joining. If your use case overlaps with any of these partners, collaboration pathways exist.

What Can Cosmos 3 Actually Do for Real-World Physical AI Projects?

Cosmos 3 is designed to be used at multiple stages of a physical AI development pipeline, not just as a single-purpose tool. According to NVIDIA, developers can use it in three distinct roles:

• Vision Language Model (VLM) — Use Cosmos 3 as a pure understanding model that processes video and text to answer questions about physical scenes, identify objects, and reason about motion and spatial relationships. This is useful for inspection systems, monitoring pipelines, and any application where understanding what is happening matters more than generating new content.
• World Model — Generate future observations given a current scene and an action. Feed in a robot's current camera view and a proposed action sequence; Cosmos 3 predicts what the robot would see after taking those actions. This is the foundation of model-based reinforcement learning and sim-to-real transfer — you can evaluate millions of action hypotheses in simulation before deploying a single one in the real world.
• Action Model — Generate action sequences directly from input observations. The model produces the control signals a robot or autonomous vehicle should execute based on its current sensory input and a goal. This is the end-to-end case: raw input in, motor commands out.

One of the most practically significant capabilities is synthetic data generation for rare events. Real-world robot datasets are expensive because collecting edge cases — the scenarios where robots fail, where unexpected objects appear, where conditions degrade — requires either running many real-world trials or paying human operators to create them. Cosmos 3 can generate those scenarios computationally, which is why NVIDIA describes training cycle compression from months to days as one of the headline outcomes.

For teams evaluating their model deployment options, our AI model comparison tool now tracks physical AI models alongside standard language models. You can also track which open source models are still fully free versus those that have added access restrictions in our free tier tracker.

The launch was announced at NVIDIA GTC Taipei, reflecting NVIDIA's continued push beyond hardware into the AI software layer. With Cosmos 3, the company is positioning itself as the infrastructure provider not just for training AI, but for the physical world models that will power the next generation of intelligent machines. Whether that bet pays off depends on how quickly the developer community adopts Cosmos 3 as a base — and the coalition structure, combined with the free license, is clearly designed to accelerate that adoption.