Google DeepMind reveals Genie 4 with real-time audio and steering for virtual worlds

Google DeepMind has introduced Genie 4, its latest world model designed to generate interactive simulations frame by frame from natural language prompts. The system is currently available as a research preview to trusted testers, with plans to roll it out to Google AI Ultra subscribers and early access API partners in the coming weeks. Compared to its predecessor, Genie 4 doubles simulation length to roughly two minutes and adds synchronized real-time audio such as ambient noise, dialogue, and music. [1][3][4]

The updated model introduces live steering, letting users alter weather, camera angles, or objects on the fly mid-simulation. It works alongside Google's Gemini models, which help direct scene actions and check frames to ensure more believable human characters and physics. However, Google DeepMind notes that limitations remain: objects and scene details still drift during longer runs, in-world text can be illegible, and audio accuracy is not fully solved. [1][4]

Google is pitching Genie 4 beyond video games, targeting robotics simulation, interactive learning, and rapid storyboarding for filmmakers. The announcement arrives amid a broader race to commercialize world models, with competitors like Runway releasing playable environments while hardware and automotive giants snap up simulation startups. [1][3][4]