Google DeepMind Unveils Genie 4 Interactive Real-Time World Model

Google DeepMind has introduced Genie 4, the successor to the Genie 3 world model demonstrated in August 2025. Unlike conventional video generators that output complete files, Genie 4 calculates scenes frame by frame in real time, responding continuously to user inputs via keyboard controls and text prompts.

For the first time in DeepMind's real-time world models, Genie 4 natively generates synchronized background audio, such as wind moving through trees, alongside responsive character dialogue. Characters in the simulation are designed to speak with one another, react to environmental changes, move naturally, and be directly controlled by users.

Powered by Gemini Models in the Background

Genie 4 builds on capabilities from DeepMind's video model, Gemini Omni, to achieve realistic visuals and behaviors. To maintain environmental consistency across scene changes, underlying Gemini models manage event planning in the background. If a user enters a prompt to change the weather, Gemini adjusts the instruction to match the running environment seamlessly.

In multi-step scenarios, Gemini 3.8 Flash continuously monitors the visual feed, breaks objectives down into individual steps, and uses new frames to verify whether each action succeeded, revising its plan if conditions shift.

DeepMind is also pairing Genie 4 with its robotics initiative, Gemini Robotics, using Gemini 4 Argon. In one demonstration, Argon divided a checkers game into step-by-step instructions. Genie 4 then simulated Apptronik's humanoid robot Apollo playing checkers based on those instructions, conducting the entire process virtually without requiring a physical robot.

Current Limitations and Access Plans

DeepMind emphasizes that Genie 4 remains an experimental research preview. While the model maintains structural coherence over brief interactions, visual details can shift and persistent objects may vanish during extended sessions or after multiple prompt modifications. Research is ongoing regarding accurate real-time sound, complex physics, and rendering readable text.

All media produced by Genie 4 will incorporate Google's invisible SynthID watermark to flag AI-generated content.

At present, access to Genie 4 is restricted to select testers. DeepMind plans to roll out the model soon to Google AI Ultra subscribers through the Project Genie web application, as well as to partners with early access to the Gemini API and Google AI Studio. In the long term, DeepMind envisions hybrid world models that combine language and code reasoning with pixel-level simulation as key training grounds for AI agents and physical robots.