Genie 3

Google DeepMind's 2025 real-time interactive world model: one text prompt generates a virtual world you can walk through.

Genie 3 is the general-purpose world model Google DeepMind released on August 5, 2025. Given a text description, it generates a dynamic world that can be explored and controlled in real time: 720p at 24 frames per second, holding consistency for several minutes, with visual memory reaching back about 1 minute. Frames are generated autoregressively one at a time, each conditioned on a growing history of past frames and the user's actions; it also supports “promptable world events,” letting text change the weather or add new objects or characters. This lets it serve as an environment for training and evaluating agents without hand-building a simulator for every task. Its limits are a restricted set of actions an agent can take and sessions of only a few minutes. At launch it was available only to a small number of researchers and creators; on January 29, 2026, Google opened it to AI Ultra subscribers as Project Genie (capped at 60 seconds per session), and Waymo has also built a self-driving world model on top of it.

ExampleDeepMind placed its SIMA agent inside a world generated by Genie 3 and gave it a goal. The agent only sends navigation actions like move forward or turn; Genie 3 has no idea what the goal is and simply simulates the next frames from those actions, and as long as consistency holds long enough, the agent can carry out longer action sequences.

Also called

Genie3

Related

World Model · Interactive World Model · Genie (Original) · Genie 2 · Waymo World Model · Autoregressive Video Generation

Sources

Genie 3: A new frontier for world models (Google DeepMind blog)
Genie (world model) - Wikipedia

As of

2026-01