Genie 3: A New Frontier for World Models

Key Claims (vendor source — technical report)

• 11B-parameter autoregressive transformer that generates dynamic interactive worlds from text prompts
• Real-time interaction at 24 fps, 720p resolution — first world model to offer real-time control
• Persistence — environments stay largely consistent for "a few minutes"; visual memory extends approximately one minute backward
• Promptable world events — users can inject text commands mid-exploration to alter the generated world
• Emergent physics — water, lighting, and ecosystem dynamics emerge from video pre-training without explicit physics engine

Generation Capabilities

• Dynamic physics simulations (water, lighting, environmental interactions)
• Natural ecosystems with wildlife and vegetation
• Historical and geographical locations (e.g., Ancient Rome reconstructions)
• Animated and fantastical scenarios
• Real-time responsive environments

Comparison to Genie 2

Advances over Genie 2 in control, resolution, and interaction latency. Genie 2 was non-real-time; Genie 3 is the first version to support real-time navigation.

Acknowledged Limitations

• Constrained agent action space
• Difficulty modeling complex multi-agent interactions
• Imperfect geographic accuracy for real locations
• Poor text rendering unless specified in prompts
• Limited to minutes rather than hours of interaction

Access & Release

Launched February 19, 2026 as "Project Genie" — limited research preview for Google AI Ultra subscribers in the U.S. (18+). Plans to expand to additional testers.

Why This Matters (Editorial)

Genie 3 is the strongest data point for the generative camp's counter-thesis to JEPA: that pixel-space prediction, if scaled enough, is a viable path to world models. LeCun argues pixel prediction is computationally intractable; Genie 3 argues it works well enough for interactive simulation at minutes-scale horizons. The empirical test over the next 12-24 months is whether generative world models can scale to hours-long horizons and robotic control, or whether JEPA-style abstract prediction pulls ahead on those fronts.

Source: Genie 3: A New Frontier for World Models by Google DeepMind