Google DeepMind is rolling out Genie 3, a new AI model that generates 3D worlds users and AI agents can explore and interact with in real time. This latest version boosts session lengths from mere seconds in Genie 2 to several minutes of continuous interaction.
Genie 3 remembers where items are in the environment. If you look away and then back, things like wall paint or chalkboard writing stay put. The generated worlds come in 720p at 24fps.
DeepMind also added “promptable world events.” You can change weather or drop new characters into the scene just by typing a command. 
Genie 3 launches as a limited research preview, available to a small group of academics and creators only. Google is testing risk controls and interaction limits before wider release. Text in worlds often only appears clearly if included in the initial input description.
The rollout follows last December’s Genie 2 demo, which produced only brief interactive spaces and lacked stable memory of scene details.
Google’s blog says users will get longer, more stable AI-generated worlds with Genie 3 but cautions broad access is still “explored” for the future.
Google DeepMind stated:
“Genie 3 supports a few minutes of continuous interaction, a big step up from the 10–20 seconds possible with Genie 2.”
“The model keeps spaces in visual memory for about a minute, meaning if you turn away and back, things remain consistent.”
“Promptable world events allow altering conditions like weather or adding new characters within the generated environments.”
“This launch is a limited research preview for a small cohort to better understand risks and mitigation.”
“We are exploring ways to bring Genie 3 to additional testers over time.”
