Google DeepMind Launches Genie 4 World Model With Real-Time Audio, Interactive Characters

Google DeepMind unveiled Genie 4, a world model that generates explorable, interactive virtual environments from text prompts, marking the company's push beyond video generation into real-time simulation. Simulations currently last about two minutes, with consistency degrading over longer runs. The model adds native audio generation for the first time in the Genie line — ambient sound, music and character dialogue synced to on-screen events — plus physics simulation and AI-controlled characters that respond to user input. Google integrated its Gemini models into the system, with Gemini Omni providing world understanding and Gemini 3.8 Flash planning tasks and verifying completion while Genie 4 renders the environment. Google demonstrated the setup by simulating an Apptronik Apollo humanoid playing checkers. Tests are underway in film, gaming, news, education and engineering; Genie 4 is available to select researchers, with plans to expand to Google AI Ultra subscribers via Project Genie and to partners through the Gemini API and AI Studio.