Google DeepMind’s Genie 4 adds real-time dialogue and native audio to simulated worlds

Google DeepMind has unveiled Genie 4, a real-time world model that generates native audio alongside interactive simulated environments. The company said it was its first model of this kind to produce sound directly, including ambient noise, music and dialogue that changes as users act or events unfold in a scene. Users can set events in motion and steer simulations through natural-language instructions. They can move a camera, alter the weather or introduce characters, while the model generates a corresponding world in which people move, react and speak. Google DeepMind said settings can carry contextual sounds such as wind through trees or waves on a beach. The model incorporates Gemini Omni to anchor simulations in real-world appearance and behaviour, while other Gemini models provide reasoning intended to preserve coherence and continuity. Gemini 3.8 Flash can observe scenes, break goals into steps and review frames in real time, according to Google DeepMind. The company said this combination enables reliable causality, procedural behaviour and multi-step reasoning. Google DeepMind presented Genie 4 as an early experimental model with enhanced photorealism, fluid movement and simulations of physics including aerodynamics and collisions. It cited a demonstration in which Gemini 4 Argon and Genie 4 simulated Apptronik’s Apollo humanoid playing checkers through real-time task decomposition. The company sees world simulation as a route toward AGI and as a way to build training environments for robots and agents. It also said the technology could support creative tools for filmmaking, journalism and gaming, as well as manufacturing. An earlier Genie model converted text, images, photographs and sketches into playable environments, using a learned latent-action interface for frame-by-frame control. All Genie 4-generated video and audio will carry imperceptible SynthID digital watermarks, intended to help identify AI-generated content. Google has also said that the system can struggle with visual consistency and audio accuracy in longer, more complex simulations; Gizmodo reported that its release includes safety evaluations and guardrails intended to block harmful or malicious content. The initial release is a research preview for trusted testers, including video-game developers experimenting with natural-language prototypes. Google plans access for Google AI Ultra subscribers through Project Genie, its prototype for creating and exploring real-time worlds, and for early-access partners through the Gemini API and Google AI Studio. Project Genie can build navigable environments from text, images or Google Street View, allowing users to walk, ride, fly or drive through them.