Google DeepMind unveils Genie 4: a world with native audio and live dialogue

Google DeepMind has announced Genie 4, the next generation of its world model. The main new feature: the system builds an interactive world in real time, complete with sound, and the characters in it talk and react to the user's actions. The announcement was written by Raia Hadsell, Diego Rivas and Hang Qi. For now, Genie 4 is available as a research preview to a small group of trusted testers, but the roadmap is already clear: first Google AI Ultra subscribers in Project Genie and early Gemini API and Google AI Studio partners, then a broader release.

Gemini inside the simulator

Genie 4 is natively multimodal. Music, background noise and interactive dialogue are generated together with the world rather than layered on top of finished video. According to DeepMind, all the sounds in the demo clips were produced by the model itself.

Architecturally, Genie 4 draws on the capabilities of the Gemini family. Gemini Omni ties the simulation to the real-world look and behavior of objects. The other Gemini models handle the complex reasoning that keeps the world coherent and continuous. For a world built on the fly, this is the weakest point: once the model loses the thread, a platform changes shape and the hero turns into a different person. Gemini acts as a director here: between user inputs, it watches the scene and nudges it forward so the world doesn't freeze.

Controls remain simple. Navigation and camera work from an ordinary keyboard, and the scene can be changed with text: switch the weather, move the camera, add a person or an object. Gemini adapts such a command to the context so the new element doesn't look like it was pasted in from another clip.

Physics, people and sound

DeepMind calls the simulation of people the most noticeable leap: characters move naturally, react organically to their surroundings, and can be controlled in real time. Real-world physics is claimed separately: aerodynamics, collisions and other scenarios.

The demos are chosen to test different sides of the model:

• a saxophonist in a baggy brown suit on a hovering concrete platform in the middle of a lake, styled like documentary 16 mm film, with realistic instrument sound and wind noise;
• an endless ribbon of pink satin winding in S-shaped curves across a hilly meadow with wildflowers;
• a first-person view from the wicker basket of a hot air balloon over Antarctic glaciers, with other balloons drifting on the horizon;
• a watermelon on a stone pedestal in a courtyard: a slow-motion clip with splashes lit by backlight;
• a macro shot of a maple seed falling from a tree;
• a crowded rural Japanese station in the 1980s, with a long dark-red train on the nearest track and an inspector in a dark blue uniform;
• a fully three-dimensional world made of stained glass, where even the water, sky and grass remain glass and crystal.

There is also a separate comparison clip of Genie 3 and Genie 4, in which the difference in photorealism, smoothness of motion and sound is plainly visible. The prompts in the demos have two layers: a description of the environment and a description of the character. Some of them are very detailed and specify composition and lighting, as well as prohibitions such as "no HDR" or "no empty platforms."

Why DeepMind is doing this

The company states its goal this way: understanding and simulating the world is a key part of the path to AGI. The practical side is also stated plainly. Safe, endlessly variable simulations are needed as training environments for robots and agents, and they also open the way to new interactive formats.

For robotics, the value is obvious. Real-world test sites are expensive and limited, while a world model with plausible physics and human behavior can produce the scenes you need from a text description. The dialogue layer gives agents an environment where others talk to them and react to them.

For the game industry and content creators, Genie 4 remains a showcase for now. There is no access, no pricing and no usage limits, and the clips were recorded under controlled conditions. If you work with agents or simulations, it is worth getting in the queue for Project Genie and seeing how well the world holds together over long stretches. Developers should keep an eye on the Gemini API and Google AI Studio: that is where early partner access will appear. The main question for the coming months: can agents be trained in such worlds in a way that lets their skills transfer to reality?