Google has announced the launch of Genie 4, a multimodal neural network capable of creating realistic interactive simulations and instantly responding to user commands. The new development is a step towards creating universal artificial intelligence and allows generating not only visual sequences, but also object physics and spatial sound.
Google has announced the launch of Genie 4, a multimodal neural network capable of creating realistic interactive simulations and instantly responding to user commands. The new development is a step towards creating universal artificial intelligence and allows generating not only visual sequences, but also object physics and spatial sound.
According to the announcement , Genie 4 is closely integrated with the Gemini ecosystem, where the latter acts as a coordinator responsible for the logic and continuity of events in the generated world.
Users can control the simulation using standard keyboard commands or text prompts. The neural network allows you to change the weather, move the camera, or add new characters on the fly, while the system automatically adapts these requests to the general context of the scene. Particular attention was paid to the calculation of physics: the model simulates aerodynamics, collisions of objects, and provides natural movements of people in the frame.
The developers call native audio generation a significant innovation. Genie 4 independently creates background noises, music, and generates human speech. Characters in the generated world can have interactive dialogues and react to user actions in real time.
To ensure the sequence of complex events, the system uses the Gemini 3.8 Flash model as a kind of management agent. It continuously analyzes frames, breaks down global goals into small steps and adjusts the action plan if the user intervenes in the process. In parallel, the company is testing integration with Gemini 4 Argon to create training environments in robotics.
The tool is currently being tested across several industries. Game developers are using it to turn ideas into interactive prototypes without writing code, filmmakers are testing different angles and lighting in a quick “draft” mode, and photojournalists are experimenting with turning flat archival footage into dynamic spatial locations. The project team says that combining the systems is a step toward creating hybrid models that are “far more coherent and manageable than either approach could achieve alone.”
At the same time, Google acknowledges that the model has certain limitations. During long sessions or after a large number of edits, objects in the simulation may disappear or deform, and the accuracy of text display needs improvement. To prevent the spread of fakes, all generated content is automatically marked with a digital SynthID watermark.
Genie 4 is currently in a closed research preview for trusted partners. It will eventually be available to Google AI Ultra subscribers as part of Project Genie, as well as to developers through the Gemini API and Google AI Studio.
