World Model
also called Interactive world simulation
A model that simulates a place you can move through, instead of a clip you watch.
A generative model that predicts how a simulated environment looks, sounds and behaves frame by frame as a user acts in it, so the world responds to input rather than playing back a fixed sequence. Google DeepMind’s Genie 4, announced October 27, 2026, is the current reference point: a text or image prompt becomes a navigable world with audio generated natively alongside the visuals, steered live with further prompts.
The distinction from a video model is interactivity. A video model produces a clip and you watch it; a world model produces a place and you move through it, changing the weather, the camera or the cast mid-scene and having the simulation adapt instead of restarting. That is why world models are pitched as training environments for robots and agents as well as a medium for interactive experiences.
The October 2026 state of the art is a research preview, not a product: Genie 4 is available only to trusted testers, with Google AI Ultra subscribers and Gemini API partners next in line. DeepMind names the limits plainly — drift over long sessions, complex physics, audio accuracy, legible text and spatial accuracy — and coverage reports scenes holding together for about two minutes, roughly double Genie 3.
The architecture detail worth noting is the second model in the loop. Coverage of Genie 4 reports Gemini 3.8 Flash acting as a guiding agent: watching rendered frames, breaking a goal into steps and checking each one landed, while Gemini Omni anchors how things look and behave. The pattern — plan, act, check — is the same loop a well-built agent uses, applied to generation itself rather than to tool calls.
A training world, not a video
A robotics team needs varied environments to train an agent, and hand-building each one is the bottleneck.
A world model generates the environment from a prompt and lets the agent act in it, so the team gets a new training scene per prompt instead of per modelling sprint. The catch is the drift DeepMind discloses: over a long session the world stops matching its own rules, so training runs need to be short or regularly re-grounded.
Calling any generated video a world model. If you cannot act in it and have it respond, it is a video, not a world.
- blogGenie 4: A new generation of world simulation ↗
DeepMind’s own announcement: Genie 4 as a natively multimodal real-time world model in research preview for trusted testers, with Gemini anchoring realism and guiding the simulation. The two-minute consistency figure is from coverage, not this post.