DeepMind's Genie 3 generates navigable, real-time interactive worlds
The system renders explorable 720p scenes at 24fps from a text prompt, holding roughly a minute of visual memory, and was released only to a small research cohort.
- Models & capabilities
- Notable
Google DeepMind introduced Genie 3, a “world model” that generates an explorable environment from a text prompt and renders it in real time as a user navigates through it. DeepMind reported 720p resolution at 24 frames per second, with the generated world staying visually consistent for several minutes and holding roughly a minute of memory of what the user had already seen — enough, for instance, for an object moved out of view to still be there on return.
DeepMind was explicit that Genie 3 was not a product but a research tool aimed at training AI agents. The team tested it with SIMA, DeepMind’s generalist game-playing agent, letting it pursue instructions inside Genie-generated scenes rather than pre-built game environments, and argued that world models more broadly are “a key stepping stone” toward AGI because they let agents be trained across an effectively unlimited range of simulated environments rather than a fixed set of games or robotics testbeds.
Access was limited to a “small cohort of academics and creators” as a research preview, with DeepMind saying it was exploring wider access later. The release followed Genie 1 (2024) and Genie 2, and its emphasis on real-time interactivity and memory — rather than just longer or higher-resolution generated video — distinguished it from the video-generation models (Sora, Veo) that dominated generative-media coverage that year, positioning world models as a separate research track aimed at agent training rather than content production.
The distinction mattered to how the release was read. A video model that generates a fixed clip from a prompt cannot be steered mid-generation or used to train an agent that needs to act and observe the consequences; a world model that renders in response to a user’s or agent’s actions, holding the scene consistent as they move through it, is a different kind of artefact aimed at a different problem — simulated environments cheap enough, and varied enough, to substitute for the hand-built game levels and robotics testbeds that had constrained agent training until then.