Google DeepMind has introduced Genie 4, a research world model that generates interactive scenes with natural-language control and native audio. Access is limited to trusted testers for now.
Google DeepMind has introduced Genie 4, a research model designed to create interactive simulated environments that users can steer in real time. Announced on October 27, 2026, the model builds on the Gemini family and is intended to maintain a scene as a user changes it through natural-language instructions.
Google’s examples describe users starting a scene and then asking the system to move the camera, change the weather or introduce new characters. The model also generates audio alongside the visuals, including ambient sound, music and interactive dialogue. That combination makes the work relevant not only to AI research but potentially to creative tools, games, filmmaking and simulated training environments.
How Genie 4 is described to work
Google says Gemini Omni helps anchor simulations in real-world appearance and behaviour, while other Gemini models provide the reasoning needed to keep the scene coherent as it evolves. In other words, the model is not only generating a static picture or a short clip; it is intended to maintain a world that responds to ongoing instructions.
World models are an active research area because they could provide environments where robots and AI agents can practise actions before interacting with the physical world. They may also give creators a more flexible way to prototype scenes and interactive experiences. However, the reliability of long-running simulations, consistency of objects and physical behaviour all remain important challenges.
Who can try it?
Google says Genie 4 has begun rolling out to a limited group of trusted testers. It is expected to become available to Google AI Ultra subscribers through Project Genie, as well as early-access partners using the Gemini API and AI Studio. That is a controlled research rollout, not a fully open public release.
The announcement is best treated as a research milestone. The practical impact will depend on how stable and controllable the generated environments become, how much compute they require and whether developers can build useful applications on top of them.
