Google DeepMind opened Genie 4 to a small group of trusted testers on October 27. The new world model holds a scene together for two minutes, takes typed directions mid-simulation, and, for the first time in the line, makes its own sound.
Genie 4 Is a World Model You Can Steer While It Runs
Genie 4 doesn't answer questions. It builds a small world that keeps running, and you move through it with keyboard controls and typed sentences.
That's the whole idea behind a world model: a system that predicts what happens next in an environment, frame by frame, instead of pulling text it has already seen. Google DeepMind calls Genie 4 its most realistic interactive model so far, and it's shipping as a research preview rather than a finished product.
The controls are the part to notice. Mid-simulation, you can ask for rain, move the camera, or drop a new person into the scene. Gemini reads the request, adapts it to what's already on screen, and keeps the world progressing between your inputs. Type something that doesn't fit the setting and the model rewrites the instruction before it acts on it.
Google DeepMind says the physical behavior improved too. Aerodynamics, collisions, and the way people walk and react all sit closer to the real thing than they did in Genie 3, according to the company.
Something changed in the background as well. Gemini models now anchor the simulation's look and its reasoning. Gemini Omni handles how the world appears and behaves, while other Gemini models work on keeping the scene coherent as it stretches out.
Genie 4's Native Audio Covers Music, Noise and Dialogue
This is the first real-time world model from Google that generates audio natively, and the company treats that as the headline change. Wind in trees, waves on a beach, a radio in the corner: each world carries the sound of its own setting.
Speech is the harder claim. Characters in a Genie 4 scene can talk, react to what you type, and adjust mid-conversation. Google DeepMind says the dialogue is generated live rather than stitched on afterward.
Anyone who has watched AI video tools knows why this matters. A silent clip looks impressive for three seconds and fake for thirty. Sound is one of the cues your brain uses to decide whether a scene is real, and getting it wrong is worse than leaving it out.
The company is candid about the limits. Audio accuracy is still an open research problem, and the team says it hasn't fully solved it. Music and ambient noise land better than speech, which is the part testers are most likely to notice.
Two Minutes of Scene Consistency, and Where It Drifts
Genie 3 held a scene together for about a minute. Genie 4 doubles that to two minutes, which sounds modest until you try to keep a room's furniture in place across a long camera move.
Longer runs expose the seams. Details and objects drift the further a simulation goes, and text inside the world still isn't reliably legible. Diego Rivas, an engineer on the team, said during a briefing that Genie 4 is early and that nobody has settled what the end-to-end product should feel like.
That kind of admission is worth more than a benchmark table. A world model that breaks after ninety seconds is a research demo. One that holds for ten minutes without weirdness is a tool.
Two minutes also narrows the use cases. A film scene, a robot rehearsal, a lesson inside a historical painting: all of those can fit. A persistent game world where you leave and come back tomorrow? Not yet.
Gemini Directs, Genie 4 Renders
The division of labor is the most interesting engineering detail. Gemini 3.8 Flash sits in the loop as a kind of director. It watches frames, breaks a goal into steps, and checks whether each step actually happened. Genie 4 draws what comes next.
In one experiment tied to Gemini Robotics, a Gemini 4 Argon model took a task, split it into captions, and fed them to Genie 4 while the simulation ran. The demo shows an Apptronik Apollo humanoid working through a game of checkers.
The point isn't checkers. It's the loop: plan in one model, rehearse in a simulation, then move to hardware. Robot teams already do this with hand-built physics engines. A generative one could cover situations nobody bothered to model, which is where the cost of robot testing actually lives.
Tejas Kulkarni, a research director at Google, described the direction as hybrid world models, with Gemini writing code and text while Genie absorbs it in real time and simulates at different levels of abstraction.
Project Genie and the Gemini API Come Next
Access is narrow for now. Trusted testers have it, and Google DeepMind says it will reach Google AI Ultra subscribers inside Project Genie in the coming weeks. Early access partners on the Gemini API and Google AI Studio are on the same schedule.
Project Genie is the Labs app that turns a text or image prompt into an explorable world. Genie 3 has been running there, so Genie 4 is an upgrade path rather than a new product. There's a waitlist for the Project Genie version.
Google also sells Veo for high-definition video with synchronized audio, and it points to Gemini Omni as a way to turn a rough simulation into a finished asset. Rivas said that last step is still some distance off, which keeps Genie 4 in the drafting stage for now.
Filmmakers in the tester group found that useful. They described using Genie 4 as a fast pre-visualization pass: step inside a scene, run several takes, change the light and the blocking while it's live. Rivas said the value shows up most in ideation and storyboarding, when you want to keep a flow going instead of waiting on a render.
The World Model Industry Genie 4 Enters
Google isn't shipping into an empty room. Runway released its second world model, GWM Worlds 2, about two months ago. AMD agreed to buy World Labs for $8.2 billion. Amazon has been reported in talks to buy Decart for around $7 billion, and Rivian bought Third Dimension to bring driving simulation in house.
Read those deals together and a pattern shows. World models are being absorbed by chip companies, cloud providers, and carmakers. That tends to happen right before a technology turns into infrastructure that nobody notices anymore.
The commercial pitches are the usual three: games, interactive learning, and robotics. The most concrete one is the least glamorous. Simulation is how you test a machine that costs six figures and falls over when it's wrong.
For a normal user, none of this changes a Tuesday. Genie 4 is a preview, and the consumer version arrives through a Google subscription first. The thing to watch is whether two minutes becomes ten, and whether the audio stops being a demo trick and becomes something you'd leave switched on.
