Google DeepMind on 5 Aug 2025 unveiled Genie 3, its latest “world model” capable of generating fully interactive 3D environments from a single text prompt. The third-generation system renders scenes in real time, allowing users to navigate or control on-screen agents as each frame is created. Genie 3 outputs 720-p video at 24 frames per second and can sustain coherent simulations for several minutes—an order-of-magnitude increase over the 10- to 20-second horizon of last year’s Genie 2. The model retains about a minute of visual memory, enabling objects to remain consistent when they re-enter view, and supports “promptable world events” that let users alter weather, insert characters or adjust terrain mid-session. DeepMind positions the technology as a way to provide near-limitless synthetic training data for robotics, autonomous vehicles and other AI agents. “Genie 3 is the first real-time interactive general-purpose world model,” research director Shlomi Fruchter told reporters, calling the advance a significant step toward artificial general intelligence. The company is releasing the model as a restricted research preview to a small group of academics and creators while it studies safety, performance and potential commercial applications.