Google DeepMind unveils Genie 4 model for generating virtual worlds

Users can navigate the generated world using a keyboard and modify the scene with text prompts: repositioning the camera, switching the weather, and adding people and objects. According to DeepMind, among the company's world models, Genie 4 is the first to generate sound on its own—from wind and surf to music—while characters talk and react to user actions. The developers also claim more realistic collisions, aerodynamics, and human movements.

Gemini 3.8 Flash monitors the coherence of what happens: the model analyzes frames, breaks the goal down into steps, and checks their execution. Separately, the company is testing a combination of Genie 4 with Gemini 4 Argon for robotics research—in a demonstration, the virtual humanoid robot Apollo played checkers.

As 3DNews reports, citing The Verge, interaction with the world now lasts up to two minutes—twice as long as with Genie 3, which was made available to AI Ultra subscribers in the United States in January 2026. The company acknowledges limitations: during long simulations, objects may change appearance and position, while audio accuracy, physics, spatial coherence, and the rendering of readable text still require refinement.

Google sees the model's application in game prototyping, pre-visualization of scenes for cinema, and creating training environments for robots and AI agents. Currently, Genie 4 operates in a research mode with a limited group of testers. Later, access is planned to be granted to AI Ultra subscribers in the Project Genie service, as well as partners working with the Gemini API and Google AI Studio. The company has not provided a timeline.