Google DeepMind Unveils Genie 4: Real-Time World Model With Audio
3 min readGoogle DeepMind has introduced Genie 4, a new world model that builds interactive, playable environments in real time and, for the first time, generates its own sound. The Genie 4 release, announced by DeepMind researchers Raia Hadsell, Diego Rivas and Hang Qi, is now in a research preview for trusted testers, with wider access planned for Google AI Ultra subscribers.
From video generator to world simulator
A world model is an AI system that predicts how an environment changes as you act in it, rather than producing a fixed video clip. Earlier Genie versions turned prompts into explorable scenes you could walk through with keyboard controls. Those scenes were silent, and objects tended to drift or vanish the longer you stayed.
Genie 4 builds on the Gemini family. According to DeepMind’s announcement, Gemini Omni anchors the look and real-world behavior of each scene, while other Gemini models handle the reasoning that keeps everything coherent.
What Genie 4 can do
Users steer the world with standard keyboard and camera controls, and can also reshape it with text prompts: change the weather, move the camera, or drop in new people and objects. Gemini adapts each change to fit the scene. The model simulates physics such as aerodynamics and collisions, and its characters move naturally and react to what is around them.
The headline feature is audio. DeepMind calls Genie 4 the first real-time world model to generate sound natively, including ambient noise, music-like audio and dialogue that responds to inputs and events. A Gemini 3.8 Flash agent can also sit inside the loop, observing the scene, breaking a goal into steps, checking each one and replanning when the user intervenes. In one experiment, DeepMind paired Genie 4 with Gemini 4 Argon to simulate Apptronik’s Apollo humanoid playing checkers.
DeepMind is open about the gaps. Visual details and object permanence can still drift in long sessions, and complex physical interactions, legible text and long-term consistency remain active research areas. All generated video and audio carries a SynthID watermark.
Why Genie 4 matters
World models are becoming a central bet for the major labs, both as creative tools and as training grounds for robots and agents. Early testers are using Genie 4 for film pre-visualization, turning a single archival photo into a walkable environment, prototyping game concepts and generating enterprise scenarios. DeepMind frames simulation of the world as a crucial step toward general intelligence.
Access is staged. Trusted testers come first, followed by AI Ultra subscribers through Project Genie and early partners in the Gemini API and Google AI Studio. Watch for the first independent tests of how long a Genie 4 session stays coherent, and for how quickly rivals answer with audio of their own.