Google DeepMind Unveils Genie 4 Real-Time World Model With Native Audio

Google DeepMind has announced Genie 4, an experimental AI model designed to generate and run interactive worlds in real time. The system produces more than a conventional video: people can move through a generated environment, change what happens with natural-language instructions and hear audio produced as the simulation develops.

The announcement, published on October 27, says Genie 4 combines capabilities from Google’s Gemini model family with a world simulator. Gemini Omni is used to help anchor the appearance and behavior of scenes, while other Gemini models provide reasoning intended to preserve continuity as the world changes.

Scenes can be changed with natural language

A user can steer an active simulation by asking it to move the camera, change the weather or introduce a person or object. Google DeepMind says Gemini adapts those instructions to the surrounding scene and can continue guiding events between user inputs. The model is therefore intended to generate a world frame by frame and respond as the user interacts, rather than render a fixed clip from beginning to end.

The company also highlights more realistic movement and physical behavior, including scenarios involving collisions and aerodynamics. Its demonstrations include characters that move and respond to their environment, as well as simulations of multi-step tasks. In one example described by DeepMind, Gemini 4 Argon and Genie 4 simulate Apptronik’s Apollo humanoid playing checkers without using a physical robot.

Audio is generated inside the simulation

Genie 4 is presented as Google DeepMind’s first real-time world model to generate audio natively. Its simulated environments can include ambient sounds such as wind or waves, along with music and interactive dialogue. Characters can converse and react to user inputs or events, according to the company.

That capability broadens the model’s potential uses beyond visual game-like environments. Google DeepMind says trusted testers are exploring applications in creative work, journalism, education, gaming and engineering, including scene pre-visualisation, interactive prototypes and simulation-based workflows. The company also describes world models as potentially useful for creating safer training environments for robots and AI agents. Those are stated research directions, not announced commercial products.

Early access comes with limitations

Genie 4 is not broadly available. Google DeepMind has begun a research preview with a limited group of trusted testers. It says access is planned for Google AI Ultra subscribers through Project Genie and for early-access partners using the Gemini API and Google AI Studio.

The model remains experimental. DeepMind says visual details and object permanence can drift during longer sessions, after several complex prompt changes or while navigating a scene. Complex physical interactions, accurate real-time audio, legible text, spatial accuracy and long-term consistency remain active research problems. The Decoder independently reported the same restricted rollout and described the model as limited in longer sessions and complex physics.

Google DeepMind says Genie 4-generated video and audio will carry SynthID, its imperceptible watermarking technology for identifying AI-generated content. For now, the release is best understood as a controlled research preview showing how language reasoning, visual simulation and sound might operate together in an interactive environment.

Sources

https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-genie-4/

https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/

https://the-decoder.com/google-deepminds-genie-4-puts-a-modern-spin-on-agent-designs-that-predate-llms/