Runway Joins NVIDIA's Cosmos 3 Coalition to Build Open World Models for Robots
Runway joins NVIDIA's newly launched Cosmos Coalition as a founding member, a global push to build and open-source frontier world models for physical AI.
Runway, the company best known for generative video tools used on Hollywood productions, is making a deliberate move into the physical world. The company has signed on as a founding member of the Cosmos Coalition, a new global initiative spearheaded by NVIDIA to build and open-source frontier world models for physical AI. NVIDIA announced the coalition at GTC Taipei alongside its new Cosmos 3 foundation model.
A world model is an AI system that learns an internal simulation of how the physical world behaves, so it can predict what happens next and plan actions, rather than just classify pixels or generate a video clip. That predictive, action-aware behavior separates these systems from ordinary video generators and supplies the missing ingredient for robots and self-driving cars that need to reason about cause and effect.
What was actually announced
The coalition is the cooperative, open-source layer wrapped around NVIDIA's bigger product launch. Founding members include Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI. The structure is a give-and-take arrangement: members contribute models, research, and evaluation techniques while drawing on Cosmos 3 technologies, training tools, and NVIDIA DGX Cloud infrastructure for large-scale training.
For Runway, there is a concrete deliverable. The first project from the coalition will be a base model co-developed by Runway and NVIDIA. Runway co-CEO Anastasis Germanidis framed the partnership as a natural extension of the company's existing direction: "The future of AI will be built on world models. That's been the driving conviction behind Runway's research from the beginning, and it's why we're proud to help found the Cosmos Coalition alongside NVIDIA. Building these systems in the open, with a shared infrastructure and a community of researchers contributing back, is how we get there faster."
The model underneath it all
Cosmos 3 runs on a mixture-of-transformers architecture, meaning the model routes different jobs to specialized sub-networks rather than pushing everything through one monolithic stack. It pairs a reasoning transformer with an expert generation transformer, letting the system understand object interactions, motion, and spatial-temporal relationships before generating video and action trajectories.
NVIDIA trained Cosmos 3 on 20 trillion tokens of multimodal data, including nearly a billion images, 400 million real and synthetic videos, ambient audio, text, and action data from humans and robots. The model natively understands and generates text, images, video, ambient sound, and actions, and NVIDIA claims it compresses physical AI training and evaluation cycles from months to days.
Two versions ship immediately:
- A "super" model for tasks requiring high physics accuracy, such as training robots and autonomous vehicles
- A "nano" model that generates results in fractions of a second
Why now, and what NVIDIA gets out of it
NVIDIA is pushing beyond chips into AI models and software, positioning itself as a foundational platform for physical AI development. A coalition of well-regarded labs all training on DGX Cloud and contributing back to a shared, NVIDIA-anchored ecosystem deepens the company's grip on the next computing wave the same way CUDA locked in the last one.
"The big bang of physical AI is just around the corner thanks to breakthroughs in multimodal reasoning language, vision and world models," said Jensen Huang, founder and CEO of NVIDIA. "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."
One practical advantage is data generation for situations that are too dangerous or expensive to film. Cosmos can synthesize rare scenarios, such as robot collisions or unusual road events, on demand. Synthetic training data of that kind addresses one of the biggest bottlenecks in robotics.
Who wins and what to watch
The smaller labs in the coalition gain the most immediately. Companies like Black Forest Labs, Skild AI, and Runway access large-scale compute and a frontier base model they would otherwise spend enormous sums to replicate, in exchange for sharing research back.
For Runway watchers, the interesting subtext is the company's pivot beyond filmmaking. The same generative-video research that powered its creative tools now targets robots and autonomous systems, a larger and more defensible market than entertainment alone. Cosmos Labs vice president Ming-Yu Liu noted that Runway brings deep research in generative video and world models to the shared effort.
The co-developed Runway-NVIDIA base model has been announced but not released, and the value of any open coalition depends on whether members keep contributing once the novelty fades. For anyone building in robotics or autonomy, a physics-aware foundation model with infrastructure and a growing roster of labs behind it now exists in the open.