DeepMind unveils Genie 4, a world model that adds sound

Google DeepMind on Oct. 27 announced Genie 4, a new world model that generates interactive simulations a person can move through and change as they run. The headline addition is sound. DeepMind said Genie 4 is its first real-time world model to generate audio natively, producing music, ambient noise and interactive dialogue at the same time as the picture.

The post, written by Raia Hadsell, Diego Rivas and Hang Qi, says the model is out now only as a research preview for trusted testers. DeepMind thanked Demis Hassabis, its chief executive, among others in the acknowledgements.

What it does: Users move through a scene with standard keyboard controls for navigation and camera, and type text prompts to steer it as it evolves, DeepMind said. The post says a person can change the weather, move the camera or introduce a new person or object mid-simulation, and Gemini adapts the prompt to fit the surrounding scene. Between those inputs, Gemini also monitors and steers the world to keep it progressing cohesively, according to DeepMind.

How it’s built: DeepMind said Genie 4 builds on the Gemini family. Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models supply the reasoning that keeps a simulation coherent. For real-time work, the post says, Gemini 3.8 Flash serves as a guiding agent in the loop, continuously observing the scene as Genie 4 generates it. DeepMind showed Gemini 4 Argon and Genie 4 working together to simulate Apptronik’s Apollo humanoid playing checkers.

What DeepMind claims: The company said Genie 4 brings “unprecedented realism,” simulates physics such as aerodynamics and collisions, and makes its biggest leap in how it simulates people, who move naturally and can be steered in real time. Those are DeepMind’s own descriptions, shown in recordings it published.

Who it’s for: DeepMind said trusted testers in creative industries, journalism, education, gaming and engineering have been trying it. Filmmakers use it as a “draft mode” to run multiple takes, game developers build explorable prototypes from a prompt, photojournalists test turning an archival photo into a simulation, and enterprise partners evaluate it for scenario generation.

What it can’t do yet: DeepMind calls Genie 4 “an early, experimental model.” It said the model can keep environments coherent over short sessions, but visual details and object permanence can drift over time, especially in longer runs, after multiple complex prompt edits or while navigating the scene. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency remain open research areas, the post says.

Safety: DeepMind said it ran internal safety evaluations before the preview, and that all video and audio Genie 4 generates will carry SynthID, its imperceptible watermark for AI-generated content.

Why it matters: Genie 4 moves world models from silent scenes toward simulations that look, sound and talk, and DeepMind ties that to its larger goals: the post calls understanding and simulating the world “a crucial aspect of building AGI” and says these simulations can serve as training environments for robots and agents. The checkers demo with Apollo shows that pitch in practice. But almost everything the public can judge so far comes from DeepMind’s own recordings, and the company’s list of limits is long, including drift in longer sessions and audio accuracy. A preview limited to trusted testers also means few outsiders can test the realism claims yet. The real measure will be how the model holds up once Ultra subscribers and developers can run it themselves.

What’s next: DeepMind said Genie 4 is “coming soon” to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and Google AI Studio. A waitlist is open in Project Genie.