Runway Introduces GWM-1: A Revolution in Reality Simulation // Runway's GWM-1: From Video to Robots and Avatars // New GWM-1 World Model Turns AI into an Interactive Tool // Runway's GWM-1: Simulations for Games, Robots, and Conversations // GWM-1: Runway Expands Beyond Hollywood into the Real World
Runway, a company known for creating videos using artificial intelligence, is now introducing something new. Its first family of world models is called GWM-1. This model is built on its Gen-4.5 video model and works by predicting frames one at a time. It runs in real time and can be controlled interactively, for example through camera movement, robot commands, or sound. According to Runway's official announcement, the goal is to create simulations that mimic the real world, including its physics and interactions.
GWM-1 is not just a single model, but three variants, each tailored to a different field. Each was additionally trained on specific data. According to Anastasis Germanidis, Runway's CTO, developing a great video model was the necessary first step toward creating a world model. The company believes that directly predicting pixels is the best path toward universal simulation. At sufficient scale and with the right data, such a model understands how the world works.
GWM-1 Variants: Worlds, Robotics, and Avatars
The first variant is called GWM Worlds. This model allows users to explore digital environments in real time. The user specifies a scene through a text description or image, and the model generates a world with an understanding of geometry, physics, and lighting. The simulation runs at 24 frames per second in 720p resolution. As you move, the world remains consistent—you can turn around, and what was behind you is still there. You can define the physics, such as whether an agent rides a bicycle on the ground or flies like a drone. According to Runway, it can be used for games in which players freely explore worlds without every detail having to be designed manually, or for education, such as exploring historical sites. It is also suitable for training agents that are learning to navigate the physical world.
Another variant is GWM Robotics. This model generates synthetic data for training robots and evaluating their behavior. It supports the generation of videos conditioned on robot actions, including alternative scenarios that explore different outcomes. Runway offers a Python SDK for this model, enabling the generation of videos from multiple viewpoints and in long sequences. Synthetic data expands existing datasets with new objects, task instructions, or environmental changes, such as different weather conditions or obstacles. This improves the robustness of trained policies without the expense of collecting data in the real world. The model also tests whether a robot violates rules in different situations, which is faster and safer than testing on physical robots. The company is in contact with several robotics companies regarding the use of this model.
The final variant is GWM Avatars. This model creates realistic avatars that simulate human behavior. It generates natural movements, facial expressions, eye movements, lip synchronization, and gestures while speaking and listening. It works with photorealistic or stylized characters and maintains quality even during long conversations. According to Runway, it can be used for real-time education, such as personalized tutors that respond to questions with natural expressions. Other applications include customer support, training simulations, and interactive games. This model will soon be available in Runway's web application and through its API.
Runway plans to merge these three variants into a single foundation model that would cover different domains and action spaces.
Gen-4.5 Video Model Update
In addition to GWM-1, Runway has updated its Gen-4.5 foundation model, which recently outperformed models from Google and OpenAI on the Video Arena leaderboard. It now adds native audio generation and long, multi-shot generation. Users can now create videos up to one minute long with consistent characters, dialogue, sound effects, and backgrounds. The model generates realistic dialogue, sound effects, and immersive backgrounds.
Audio editing is also coming—you can modify existing audio in videos, add dialogue, or change sounds. Another new feature is multi-shot editing, where a change in one scene is applied throughout a video of any length, such as changing a character's eye color or replacing the background with a jungle. These features are available to paying users and bring Runway closer to tools such as Kling, which also offers everything in a single package.
These advances are impressive, especially if the claims of consistency across longer sequences are confirmed. Runway has also announced a partnership with CoreWeave to use Nvidia GB300 NVL72 racks for training and inference.
Additional Details
GWM Robotics is available through an SDK, and the company is in discussions with businesses about using Robotics and Avatars. GWM Worlds is suitable for VR and immersive experiences, where endless explorable realities can be generated. According to Ivan Mehta and Rebecca Bellan of TechCrunch, GWM-1 is more general-purpose than competing models such as Google Genie-3 and can be used to train agents in robotics or the biological sciences.
Runway emphasizes that world models represent the frontier of progress in artificial intelligence because language models alone are not enough to solve complex problems such as robotics or scientific discovery. Simulations enable accelerated learning from mistakes without real-world risks. According to Cristóbal Valenzuela, CEO of Runway, GWM-1 is a major step toward universal simulation. The company offers a form for requesting early access to these tools.
