Google DeepMind has announced Genie 4, a world model designed to create interactive simulations that respond in real time to keyboard controls and natural-language instructions. The research preview combines generated visuals, sound, dialogue, motion and physical behavior inside a continuously evolving environment.
Unlike a conventional video generator that produces a fixed sequence, Genie 4 is intended to react while a user explores. DeepMind says users can navigate with standard controls and alter a scene through text—for example, changing the weather, moving the camera or introducing a person or object. Gemini models help interpret those instructions and maintain continuity between interventions.
The system is natively multimodal. It can generate ambient sounds, music and interactive dialogue alongside the visual world rather than adding them through a separate post-production step. DeepMind's demonstrations include outdoor landscapes, moving fabric, physical objects and human characters, with audio generated as part of the simulation.
Physics and movement are major areas of emphasis. The company says Genie 4 improves its handling of aerodynamics, collisions and the way people respond to their surroundings. Natural human motion is especially important for interactive environments, where small inconsistencies become more visible as a user changes direction or introduces an unexpected event.
DeepMind describes Gemini Omni as an anchor for real-world appearance and behavior, while other Gemini systems supply reasoning used to preserve coherence. Between direct user commands, the models can steer the scene so activity continues. This architecture is presented as a way to make generated worlds less static and more persistent, though the announcement does not provide independent evaluation of reliability or long-duration consistency.
The first release is limited. Genie 4 is rolling out to trusted testers as a research preview while DeepMind gathers feedback. The company plans later access for Google AI Ultra subscribers through Project Genie and for selected early partners using the Gemini API and Google AI Studio. A waitlist is available, but the announcement does not promise broad consumer or developer availability on a fixed date.
DeepMind connects the project to research on artificial general intelligence. Variable simulations could provide training environments for robots and software agents without requiring every scenario to be staged in the physical world. The same technology could support interactive entertainment, prototyping and creative tools. Those applications remain prospective and will depend on controllability, safety, cost and whether generated physics is accurate enough for a particular task.
Genie 4 is therefore best understood as a capability preview, not a finished simulation platform. The demonstrations show DeepMind's direction: generated environments that can be entered, heard and changed rather than merely watched. Testing by outside users will determine how consistently the model preserves objects, characters and causal relationships beyond the curated examples shown in the launch material.
