Genie 3 real-time world model release

TL;DR

Genie 3 is DeepMind’s newest general‑purpose world model that can generate richly detailed, interactive 3D environments in real time (24 fps, 720p) and maintain visual consistency for several minutes, opening new possibilities for embodied AI research and generative media.

What is Genie 3?

Genie 3 is a foundation world model that expands on DeepMind’s earlier Genie 1 and Genie 2 systems. It creates interactive worlds from natural‑language prompts, allowing users to navigate them at 24 frames per second in 720p resolution. Unlike prior models, Genie 3 supports real‑time control, longer‑horizon consistency, and a broader range of physical and visual phenomena.

Technical Advances

Real‑time frame generation

  • The model autoregressively generates each frame while conditioning on the entire past trajectory, enabling responsive navigation.
  • Computation is performed multiple times per second to incorporate new user inputs, achieving smooth 24 fps playback.

Long‑horizon environmental consistency

  • Genie 3 retains visual memory for up to a minute, preventing drift that typically plagues autoregressive video generation.
  • Consistency is demonstrated in examples such as a house‑painting task, a Victorian street with a teleporting portal, and a multi‑minute exploration of ancient Athens where static elements (e.g., trees) remain stable.

Physical property modeling

  • The model simulates water, lighting, lava flows, wind, and other natural phenomena with realistic dynamics.
  • Sample prompts showcase volcanic terrain navigation, hurricane‑affected sidewalks, and deep‑sea jellyfish movement.

Diverse content generation

  • Natural ecosystems – forests, glacial lakes, and marine environments with believable flora and fauna.
  • Fictional animation – stylized creatures, surreal landscapes, and dramatic world‑bending events.
  • Historical and geographic settings – accurate‑looking reconstructions of the Alps, Venice canals, the Palace of Knossos, and modern streets in Illinois.

Capabilities Demonstrated

Capability Example Prompt Observation
Physical simulation Volcanic terrain with lava pools Realistic particle effects and terrain interaction.
Ecosystem generation Glacial lake run with wildlife Cohesive animal behavior and vegetation.
Fictional animation Fluffy creature on rainbow bridge Stylized, high‑detail character animation.
Historical reconstruction Palace of Knossos in its heyday Plausible architectural detail.
Promptable world events Change weather or add objects via text Immediate world alteration without re‑rendering.

Embodied Agent Research

Genie 3 serves as a testbed for training agents such as DeepMind’s SIMA 2. Agents issue navigation actions to the model, which then renders the resulting world state. Because Genie 3 maintains consistency over longer action sequences, agents can pursue more complex, multi‑step goals, bringing us closer to training agents in open‑ended curricula.

Limitations

  • Action space – Direct agent actions are limited; many world changes still require textual prompts rather than embodied interaction.
  • Multi‑agent dynamics – Simulating interactions between several independent agents remains an open challenge.
  • Geographic fidelity – Real‑world locations are approximated and lack precise geographic accuracy.
  • Text rendering – Generated text is often legible only when explicitly included in the prompt.
  • Interaction duration – Continuous interaction is currently limited to a few minutes, not hours.

Responsible Deployment

DeepMind is releasing Genie 3 as a limited research preview to a small cohort of academics and creators. This staged rollout is intended to gather feedback on safety, misuse potential, and societal impact while the Responsible Development & Innovation team works on risk mitigation strategies.

Future Directions

  • Expanding access to broader tester groups and eventually the public.
  • Extending interaction duration and action space to support richer embodied tasks.
  • Improving geographic accuracy and multi‑agent simulation.
  • Exploring educational, training, and evaluation applications for robotics and autonomous systems.

Citation

If you reference Genie 3, please use the provided BibTeX entry:

@article{genie32025,
  title={Genie 3: A new frontier for world models},
  author={Ball, Phil and Bauer, Jakob and ... and Zubov, Vadim},
  journal={DeepMind Blog},
  year={2025},
  url={https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/}
}

Sources