Genie 4 can maintain a simulated scene for up to two minutes, and Google DeepMind has released it in research preview. The model adds real-time audio and lets users change scenes with prompts while they are running. The report also describes an experiment linking it with Gemini Robotics to simulate tasks.
What changed in the new preview?
Google DeepMind’s earlier Genie 3 was also released as a research preview, according to the source. Genie 4 is described as producing more realistic and interactive worlds, with improved physics and more humanlike characters. Its scene consistency can last for up to two minutes, double the predecessor’s limit, the report says. These details describe the model’s capabilities, not a guarantee that every generated scene will behave identically.
How can a user steer a scene?
A user can type prompts during a simulation to request changes. The report names adjustments such as shifting weather or camera angle, or adding people and objects. This kind of steering means the scene can be altered as it unfolds. The source does not specify which prompts are supported or how reliably the model handles each request.
What does real-time audio include?
Genie 4 generates sound to accompany its simulated world in real time. The source mentions music, environmental noise and interactive dialogue. This allows audio to correspond with an interactive scene as it changes. The report does not give figures for audio delay, quality or the range of supported sounds.
How does the robotics experiment work?
The source describes a research experiment associated with Gemini Robotics. In that experiment, a Gemini 4 Argon model divided tasks into steps, and Genie 4 simulated those steps in real time. The approach could let researchers test robot scenarios in simulation before carrying them out in the physical world. The report presents this as a possible research use, without claiming that it is already deployed for operational robot testing.
Which uses are being considered?
The report names gaming, interactive learning and robotics as possible commercial applications for world models. It gives an illustrative learning example in which someone enters a historical painting and speaks with its subjects. Film prototyping is another potential use raised by trusted testers, according to the report. Diego Rivas, a Google DeepMind engineer, described rapid ideation and storyboarding as settings where real-time interaction could be useful.
Genie 4 is currently described as a research preview, and the source gives no timetable for wider availability. It also does not explain who qualifies for access or what limits apply to generated scenes. The robotics experiment remains a research example in the material provided. More details on access, product plans and further testing would show how these ideas may develop.