TL;DR
Google DeepMind announced Genie 4 on 27 October, a world model that generates explorable environments in real time and, as DeepMind’s first real-time world model with native sound, produces audio alongside the picture. Users move through a scene with keyboard controls and change it with text prompts. It is a research preview for trusted testers for now, with Google AI Ultra subscribers and early access developers next.
What it does
A world model simulates an environment that responds as you act in it, rather than playing back a fixed clip. With Genie 4, a user can start an event and the model works out what follows: the scene’s response, its soundscape, and the movement and speech of the people in it. Prompts can alter the weather, move the camera or add a character mid-session.
Audio is the headline change. DeepMind’s own post describes Genie 4 as its first real-time world model to generate sound natively, from ambient noise to dialogue between characters that responds to the user. It also claims better physics, such as collisions and aerodynamics, and much more natural human movement.
Gemini models do the planning. Gemini Omni grounds how scenes look, and Gemini 3.8 Flash runs alongside the simulation, splitting a goal into steps, watching frames as they arrive and checking each step worked. DeepMind is also testing Genie 4 with its robotics research, so tasks can be rehearsed in simulated worlds before a robot attempts them.
Who is testing it
DeepMind says testers in film, journalism, education, gaming and engineering are using it as a quick drafting tool. Film-makers can run several takes of a scene live, photojournalists can turn one archive photograph into a navigable environment, and game developers can prototype before writing engine code. Google says all video and audio it produces will carry its SynthID watermark. DeepMind’s launch video is embedded above, and it also sits on our Videos page.
The company is open about the limits. Detail and object permanence drift over longer sessions or after repeated edits, and text rendering, spatial accuracy and audio accuracy are still being worked on. We looked at the gaming question in February in our piece on Google’s AI worlds.
Looking Forward
Access is narrow, so outside testers cannot yet check DeepMind’s claims. In our view, the robotics link matters more for business than the creative demos: if simulated worlds become reliable enough to train and test agents and robots, they could cut the cost of real-world trials. For UK creative and games studios, the immediate use is pre-visualisation, where speed matters more than perfect consistency.