Genie 3

2025-08-05 · Publisher : Google · Domain : Artificial Intelligence

Google DeepMind announces Genie 3, a general-purpose world model that creates interactive environments navigable in real time at 24 frames per second, in 720p, from a simple text prompt.

What changes

First Google DeepMind world model to allow real-time interaction, with improved consistency and realism compared with Genie 2.
Environments consistent for a few minutes, with a visual memory going back up to one minute.
"World events" controllable by text (promptable world events) to modify the environment during interaction.
Tested with the SIMA generalist agent for training agents in generated worlds.

Google DeepMind announces Genie 3, a general-purpose world model that creates interactive environments navigable in real time at 24 frames per second, in 720p, from a simple text prompt. ### What changes - First Google DeepMind world model to allow real-time interaction, with improved consistency and realism compared with Genie 2. - Environments consistent for a few minutes, with a visual memory going back up to one minute. - "World events" controllable by text (promptable world events) to modify the environment during interaction. - Tested with the SIMA generalist agent for training agents in generated worlds. ### Specifications - Generation: 24 frames per second, 720p resolution. - Acknowledged limitations: limited action space, modeling of interactions between agents, real-world locations not faithful, legible text only if it appears in the description, a few minutes of interaction. - Availability: research preview limited to a small group of academics and creators. ### Pricing Not applicable: no public availability. ### Benchmarks No numeric score published in the post. ### Key takeaways A research milestone for environment simulation and agent training, with no public access in 2025.

Genie 3

What changes

Links