Google DeepMind has introduced Genie 3, an advanced AI model that can create 3D video game worlds in real time. Unlike earlier versions, Genie 3 allows users and AI agents to interact with these environments for several minutes and even remembers where things are located when users look away.
The model is a major step forward in the development of AI-generated “world models,” which simulate virtual environments for use in gaming, education, and training robots. With just a prompt, Genie 3 can build a playable world that runs at 720p and 24 frames per second.
Unlike Genie 2, which only supported 10–20 seconds of interaction, Genie 3 extends that to a few minutes and introduces “promptable world events.” That means users can change weather, add characters, or create other dynamic actions simply by typing in a prompt.
Another key upgrade is visual memory: if you leave a scene and return, things like a chalkboard drawing or paint on a wall will still be there—just like in real life.
However, the model isn’t widely available yet. Google is launching it as a limited research preview for a small group of academics and creators. The company says it wants to better understand risks and limitations before releasing it more broadly. For now, user interaction remains restricted, and clear in-world text will only appear if described in the original input.
Google is investing heavily in world modeling AI, with a dedicated team led by a former co-lead of OpenAI’s Sora, another advanced video generation system.
