DeepMind Unveils Genie 4, Interactive Worlds with Audio

DeepMind launches Genie 4, an interactive world model calculated frame by frame in real time, capable of directly producing soundscapes and dialogues. The system remains experimental, relies on Gemini's planning to maintain coherence, and is available by invitation only, with a SynthID watermark announced for all generated media.

Known Limitations, Restricted Access, and Media Watermarked by SynthID

DeepMind describes Genie 4 as an initial and experimental model. According to the company, the coherence of the world is maintained during brief sessions, but some visual elements may vary, and objects may not remain present during prolonged simulations or after several complex prompt modifications. The company notes that it continues to work on managing advanced physical interactions, generating realistic real-time sound, and producing readable text. Genie 4 is offered in a research preview to selected testers only. Google also states that all videos and audio tracks generated by the system will be watermarked by SynthID, an invisible watermark for AI-generated content.

What Genie 4 Can Do in Terms of World, Characters, and Interaction

Genie 4 calculates environments frame by frame in real time and does not produce finished videos, reacting to inputs as they occur. Users navigate using the keyboard and orient the scene with text prompts. DeepMind presents this model as its first real-time world that directly generates sound, with each universe having a dedicated ambiance and supporting dialogues between characters in response to actions and events. The company highlights the main gain in realism in the rendering and behavior of characters, which are expected to move naturally, react to the environment, and be directly controllable.

World Coherence Relies on Gemini and Continuously Adjusts

For realistic appearance and behavior, Genie 4 relies on the capabilities of the Gemini Omni video model. Other variants of Gemini help plan the simulation to maintain its coherence, managing multi-step tasks and continuously adapting the world to prompts and changing conditions. For example, if a user modifies the weather via text, Gemini adjusts the instruction to the current scene. Between user inputs, Gemini also directs events so that the world continues to evolve logically.

Closed-Loop Monitoring with Gemini 3.8 Flash and Step Tasks

DeepMind describes a functioning model based on tasks broken down into steps. Gemini 3.8 Flash continuously monitors the scene, segments a goal into sub-goals, and then relies on the newly generated images to verify the execution of each step. If the state of the world changes or if a user intervenes, the plan is adjusted by the model.

Bridge to Robotics, Access Roadmap, and Range Context

DeepMind is testing Gemini 4 Argon to connect Genie to its robotics research. During a presentation, Argon segmented a game of checkers into different steps, and Genie 4 relied on these instructions to replicate a game played by the humanoid robot Apollo from Apptronik, without using a physical robot. Genie 4 follows Genie 3, which was presented in August 2025, and it is expected to soon be accessible to subscribers of Google AI Ultra via the Project Genie web app, as well as to certain partners with early access to the Gemini API and Google AI Studio. DeepMind views the combination of Gemini and Genie as a first step toward hybrid models that combine reasoning in language and code with pixel-level simulation. Mature world models could be used as training environments allowing robots and agents to learn through simulated experience.