Overview

• DeepMind positions Genie 3 as the first real-time, general-purpose world model to advance embodied agents and help pave the way toward artificial general intelligence.
• With simple text prompts, Genie 3 creates diverse 720p simulations running at 24 fps for several minutes and supports promptable world events like weather shifts or new objects.
• An emergent memory mechanism lets the model recall previous frames to maintain physical consistency across scenes without hard-coded physics.
• Genie 3 is limited to a small cohort of academic and creator testers in a research preview, and DeepMind has not announced a public release date.
• Despite its breakthroughs, the model still restricts agent actions, struggles with multi-agent interactions and only sustains a few minutes of continuous simulation as OpenAI teases GPT-5.