Ask a world model to render a room and it will give you walls, light and furniture. Ask it what the room sounds like and, until this week, the answer was nothing. Google’s DeepMind published Genie 4 on Oct. 27 as a research preview, and the first thing the demonstration does is turn the sound on.
Genie 4 is the company’s first real-time world model that generates audio natively. Ambient noise, music and dialogue are produced in step with the scene rather than laid over it afterward, so a door that closes in the frame is a door that is heard closing. Characters inside the simulation answer when the user speaks to them. The physics extends to human motion, gravity and objects breaking apart. A single scene now holds its consistency for as long as two minutes, twice what Genie 3 could manage.
The controls are the part that changes what the system is for. A user can edit the simulation while it runs, typing instructions that change the weather, add a person or cut to a different camera angle. That is closer to directing than to prompting: the model is not asked to produce a clip and then asked again, it is asked to keep a world running and to accept corrections inside it.
Two minutes sounds like a small number and is not. Most of the tasks that make simulated environments commercially interesting, from a robot learning to stack boxes to an agent working through a multi-step procedure, run longer than the few seconds a video model can hold together. Consistency is the difference between a demonstration and a test bench. At two minutes a simulation can hold a sequence of actions with consequences, and the consequences stay in the frame.
The audio may prove the more consequential addition. A world model that renders only pixels hands an agent a silent environment, which is fine for a benchmark built on images and wrong for anything meant to operate in a physical space. Glass breaking, footsteps behind a wall, a machine that starts to whine: sound arrives before the visual evidence and carries information about events the camera cannot see. Giving a simulated world a soundtrack gives an agent something to react to that is not already framed for it.
DeepMind has been explicit that the company sees this as a systems problem rather than a single model. Genie 4 generates the world. Gemini 3.8 Flash, wired in as the reasoning layer, understands the goal, observes the environment, checks whether the previous step succeeded and decides what to do next. One model holds the simulation, the other decides what to try inside it.
The most specific example the company has shown involves a humanoid robot playing checkers. Gemini 4 Argon, a higher-end model, broke the task into steps and handed them to Genie 4, which simulated Apptronik’s Apollo carrying out the whole sequence. The physical robot was not in the room, was not switched on and was not needed. What the experiment tested was whether a planner could produce instructions precise enough for a simulated body to execute, and whether the simulation could report back clearly enough for the planner to notice a failure.
That division of labor is where the commercial argument sits. Training a humanoid on hardware is expensive, slow and dangerous in the ways that matter: a robot that falls over costs money, and one that falls over near a person costs more than money. A simulation that can be edited mid-run, that produces sound, and that holds together for minutes at a time lets a team test failure modes it would never deliberately induce on a real machine.
The limits are real. Genie 4 is a research preview, not a product with a price and a service-level agreement. Two minutes is a ceiling, not a floor. The gap between a simulated body and a physical one, which engineers call the sim-to-real gap, is where most robotics promises have died, and no amount of scene consistency closes it on its own. A model that renders plausible physics is still a model of physics, and the errors it makes are the ones a policy trained inside it will learn to expect.
There is also the question of what these systems do to the market for training data. Robotics has been held back less by algorithms than by the scarcity of recorded interaction with the physical world. A world model that can generate those interactions, on demand and with labels attached, is a substitute for a resource that has no cheap alternative. That is why the frontier labs have been converging on this problem from several directions at once.
What DeepMind has shown is a system that is beginning to look less like a video generator with better physics and more like an operating environment. Genie 4 supplies the world, Gemini supplies the intent, and the loop between them runs without a robot, a warehouse or a laboratory. The two-minute clock is the constraint to watch: every minute added is another class of task that can be rehearsed before anyone builds the machine.
