Genie 3

What it is

Genie 3 is a general-purpose world model from Google DeepMind, announced on 5 August 2025 by Jack Parker-Holder and Shlomi Fruchter^[1]. A world model, in the sense the term carries in AI research, is a system that learns to simulate how an environment behaves, so that a person or an AI agent can act inside it and see the consequences. Genie 3 builds such environments from a text prompt rather than from hand-written scene code, and what it produces is meant to be entered rather than watched: a generated world can be navigated in real time at 24 frames per second, rendered at 720p^[1]. It is the first world model in the Genie series to allow real-time interaction^[1].

Holding a world together

Interactive generation is a memory problem as much as a rendering one. If the model forgets what it drew a moment ago, the world shifts behind the user's back. Genie 3's worlds keep their consistency for a few minutes, and the visual memory reaches back about a minute^[1]. That window defines what the system currently offers: a scene you can walk through and turn around in, for a span of minutes^[1].

Promptable world events

While a world is running, a user can change it in words — altering the weather, adding objects or characters. Google DeepMind calls these "promptable world events"^[1]. They are not only a novelty: because such events can stage what-if scenarios, they are presented as a way to help agents prepare for unexpected situations^[1]. One distinction is worth keeping in mind here: environment interventions in Genie 3 are not necessarily performed by the agent itself^[1]. The world can change for reasons the agent had no hand in, which is what makes an unexpected event unexpected.

Worlds as a testbed for agents

Google DeepMind used Genie 3 worlds to test a version of its SIMA agent^[1]. The agent pursued goals by sending navigation actions to the model^[1]. This is what a world model is for in an agent setting: the model supplies the environment and the agent supplies the behaviour, so the world becomes a place where an agent can be evaluated, instead of a fixed set of levels written in advance.

Stated limitations

The announced limitations are three: a limited action space, difficulty modelling multiple interacting agents, and interactions limited to a few minutes^[1]. Each names a different constraint. An action space is the set of moves an actor is allowed to make; if it is narrow, the world is somewhere you can travel through but not really manipulate. Multi-agent modelling matters because two agents changing the same world have to be kept consistent with each other. The few-minute horizon is tied to how far back the model's memory reaches.

Why it matters

The interesting part is the combination rather than any single number. Generating a world is one capability; generating one that responds at 24 frames per second while holding itself together for minutes is another; letting a running world be changed by a sentence is a third^[1]. Together they make a system that can be used the way a simulator is used — as a place to put an agent and watch what it does — without anyone building the place by hand.

Release

Genie 3 was released as a limited research preview to a small group of academics and creators^[1]. That is a deliberately narrow opening: the system is being shown and studied, not shipped as a finished product.