Overview

• Genie 3 generates fully interactive 3D environments in real time, rendering scenes at 720p resolution and 24 frames per second for both users and AI agents.
• The model extends continuous interaction from seconds to several minutes and uses visual memory to maintain spatial consistency when scenes are revisited.
• It introduces promptable world events that let researchers alter weather, add characters or modify environments via simple text commands.
• DeepMind is offering Genie 3 as a limited research preview to a small cohort of academics and creators to evaluate its capabilities and risks.
• While marking a key step toward training embodied agents and artificial general intelligence, Genie 3 still faces challenges in multi-agent dynamics, text rendering and hour-scale simulations.