Google DeepMind introduced Genie 4 on Tuesday, initially limiting its research preview to trusted testers. The world model generates interactive environments in real time rather than producing a finished video, allowing users to explore and modify scenes as they unfold. Google describes improvements in visual realism, human characters, and simulated physics over Genie 3.
Users can issue natural-language instructions to alter the weather, introduce people or objects, and adjust camera angles without restarting an experience. Generated audio includes music, ambient sound, and dialogue. Gemini 3.8 Flash helps steer the simulation by planning steps and evaluating individual frames.
The model can maintain scene consistency for approximately two minutes, twice its predecessor’s limit. Proposed applications include film previsualization, game prototyping, interactive learning, and robot training. In one robotics demonstration, Gemini 4 Argon planned moves for Apptronik’s Apollo humanoid in a simulated game of checkers, with Genie 4 generating the results.
Reliability remains a constraint: objects can shift position or behave unpredictably as simulations continue, and the physics is not always accurate. The team also acknowledges difficulties ensuring accurate audio and legible text within generated worlds. For robot testing, these inconsistencies can complicate judgments about whether a failure originates in the robot’s decision-making system or the simulated environment.
Google plans to extend access to Google AI Ultra subscribers through Project Genie in the coming weeks. Selected partners will also receive access through the Gemini API and Google AI Studio.
