Google DeepMind has announced Genie 4, the latest version of its world model, a system that generates an explorable scene frame by frame while you move through it with the keyboard. The headline change is sound. Earlier real-time Genie models were silent; this one produces ambient noise, music and spoken dialogue together with the picture. For anyone judging a scene rather than just looking at it, that is a meaningful step.
The model leans on the wider Gemini family. DeepMind says Gemini Omni keeps the look and behaviour of a scene close to the real world, while other Gemini models handle the reasoning that keeps it consistent. Users can type instructions during a session, such as changing the weather, moving the camera or adding a person or object, and Gemini rewrites the request to suit the scene. For longer tasks, Gemini 3.8 Flash sits in the loop, splitting a goal into steps, watching incoming frames and checking whether each step worked. One demonstration pairs Gemini 4 Argon with Genie 4 to simulate Apptronik's Apollo humanoid robot playing checkers.
The trusted testers DeepMind describes come from film, journalism, education, games and engineering. Filmmakers use it as a rough draft for previsualisation, rerunning a scene with different lighting or dialogue. Game developers use it to go from an idea to a walkable prototype before writing engine code. Photojournalists are testing whether a single archival photograph can be turned into a simulated place. All output carries SynthID watermarks in both the video and the audio.
Genie 4 sits in a field that is filling up. Our reviews of Genie 4 and of Tencent's HY-World 2.0 cover two different answers to the same question: whether a model can stand in for a hand-built 3D environment. DeepMind's bet is to tie the world model closely to its language models, so that planning happens in Gemini and rendering in Genie. It calls this a step towards hybrid world models.
The gaps are large. The post gives no resolution, frame rate or maximum session length, and no benchmark against rivals. DeepMind itself says details and object permanence drift as sessions run longer or after several edits, and that physics, audio accuracy, legible text and spatial accuracy are still being worked on. Access is limited to trusted testers for now, with Google AI Ultra subscribers in Project Genie and early partners on the Gemini API and Google AI Studio promised next, with no date or price. Until those numbers appear, it is a preview to watch rather than a tool to plan around.