Google DeepMind announces the real-time Genie 3 world model

What a world model is

A world model is a system that learns how an environment behaves: what happens to a scene as time passes, as an object moves, as something acts on it. Once a model can predict that, it can serve as a simulated environment — a place to try things out, or to test an agent without building a physical set.

A video generator is a close relative but not the same thing. A generated video plays forward on its own, the same way every time. An interactive world has to respond to what the user does, at whatever moment, from whatever viewpoint. Doing that live, rather than rendering afterwards, is the harder problem.

The announcement

On 5 August 2025, Google DeepMind announced Genie 3 ^[1]. The company describes it as a world model that generates navigable worlds from text in real time, at 24 frames per second and 720p resolution ^[1]. Those two numbers are what separate it from a clip you wait for: 24 frames per second is the rate at which ordinary video looks like continuous motion, and 720p is a normal high-definition image size, so the output is meant to be watched and steered as it is produced rather than rendered and then played back.

What the worlds are like

A generated world does not last indefinitely. Google DeepMind says environments stay consistent for a few minutes ^[1]. Within that window, the scene holds together: the layout does not dissolve, objects stay where they were put, and moving through the space does not scatter the picture.

The model also has visual memory reaching back as far as one minute ^[1]. That is the span over which it can recall what it has already shown, so a space seen earlier can still be returned to rather than replaced by something new.

Beyond moving around, a user can trigger promptable world events — asking for a change in the weather, for example — and have the environment respond ^[1]. This turns the world from a fixed stage into something that can be modified while it is running, which is closer to how a scenario is set up when testing an agent: change a condition, see what happens next.

Who can use it

Genie 3 did not launch as a general product. It arrived as a limited research preview for selected academics and creators ^[1]. That framing matters for reading the announcement: the claims are about what the model can do in a research setting, with access handed out selectively rather than opened to the public.

Google DeepMind also put the worlds to work internally, using them to test recent versions of its SIMA agent ^[1]. An agent is a system that takes actions in an environment in pursuit of a goal, and testing one requires environments to act in — many of them, with varied conditions. A model that produces navigable worlds from text is a way of getting those environments without hand-building each one.

Why it is worth noting

The interesting claim in Genie 3 is not the image quality by itself but the combination: text in, an explorable world out, generated live, with enough persistence to hold a scene together for minutes and enough memory to keep it coherent for about a minute. Consistency and memory are the properties that decide whether a generated world can be treated as a place rather than as a picture that keeps changing. Real-time generation and promptable events are what make it usable interactively.

For research on agents and embodied systems, an endless supply of responsive environments is a practical bottleneck. Genie 3, as described, points at that bottleneck — a world model that doubles as a testing ground — while the limited preview makes clear that, for now, it is a research instrument.