About
Genie is a family of foundation world models developed by Google DeepMind, aimed at generating an endless variety of action-controllable environments for training and evaluating generalist AI agents. The family started with Genie (2024), introduced in the paper "Genie: Generative Interactive Environments" — a generative model trained on internet videos that produces interactive 2D worlds from a single prompt image. Genie 2 (December 2024) is a large autoregressive latent-diffusion model that generates consistent 3D worlds, controllable via keyboard and mouse, for roughly 10–60 seconds at a time, demonstrating long-horizon memory, character animation, physics, lighting and particle effects. Genie 3 (August 2025) continues this direction, extending the generation horizon and the quality of the simulated worlds. The family serves as a backbone for research on embodied agents (such as SIMA) and on general AI capable of acting in rich interactive environments.