Google DeepMind on 5 August unveiled Genie 3, a foundation AI model capable of generating fully interactive 3D environments in real time from a single text prompt. The system renders worlds at 720p and 24 frames per second and can keep them visually and physically consistent for several minutes, maintaining a memory of events for about one minute. Users—or AI agents—can navigate the scenes with keyboard or controller inputs, while “promptable world events” let them alter conditions on the fly, such as changing weather or adding new objects and characters. Research director Shlomi Fruchter called Genie 3 “the first real-time interactive general-purpose world model,” highlighting its potential to train embodied AI agents. In a demonstration, DeepMind’s SIMA agent successfully completed goal-oriented tasks inside Genie-generated warehouses, underscoring the model’s value as a testing ground en route to artificial general intelligence. Compared with Genie 2, which was limited to 360p and about 20 seconds of interaction, the new model more than triples resolution and extends the playable horizon to several minutes. Genie 3 remains in a restricted research preview, available only to a small group of academics and creators while DeepMind studies safety and performance. The lab notes current limitations, including a constrained action space, imperfect real-world accuracy and the need to extend simulations beyond a few minutes. Google is evaluating broader access but has not set a public release date.