Google Releases Genie 4: Real-Time World Simulation, Now With Sound

Google today released Genie 4, DeepMind's next-generation interactive world model. The core change: native multimodality. Images, sound, and dialogue are generated in one pass, and you can change the scene in real time with text while it runs.

Four Key Capabilities

Real-time control. Use text prompts to modify or add things in the world. Change the weather, move the camera, add people or objects; Gemini uses context to adapt the instruction to the current scene. Between inputs, Gemini continuously monitors the scene to maintain coherent progression.

Native audio. Every world comes with sounds that match its environment: wind through leaves, waves crashing on the shore. Character dialogue can respond in real time to your input or events in the world. This is the first real-time world model in the Genie series to natively generate audio.

Physics and characters. It can simulate physical scenarios such as aerodynamics and collisions. Character motion shows the biggest improvement—movement is natural, reactions to the environment are organic, and it can be controlled in real time.

Gemini reasoning. Gemini 3.8 Flash acts as a guiding agent in the loop: it continuously observes the scene, breaks goals into steps, and checks progress frame by frame. When the user intervenes, it adjusts the plan on the fly. This lets the world support multi-step events and reliable causality.

Real-time visual comparison between Genie 3 and Genie 4

Underlying Architecture

Genie 4 is built on the Gemini family. Gemini Omni anchors the simulation to the appearance and behavior of the real world, while other Gemini models provide reasoning that maintains coherence. You set an event, and Genie 4 simulates what happens next: how the world responds, how it sounds, and how the people in it move, react, and speak.

Robotics Research

Google combines Genie 4 with Gemini 4 Argon for robotics research. Argon breaks user instructions into steps, Genie renders them into visuals, and adjustments are made in real time during simulation. Example: Apptronik's Apollo humanoid robot plays checkers in a simulated world.

Apollo humanoid robot playing checkers in a simulated world

What It Can Do

Feedback from trusted testers points to several clear directions:

Filmmakers use it as a previsualization draft mode, running multiple shots in real time and adjusting camera positions, lighting, and dialogue
Photojournalists use a single 2D archival photo to generate an enterable environmental simulation and experience historical spaces
Game developers use natural language to test visual styles and interaction mechanics before writing engine code
Robotics research: letting robots trial and error repeatedly in simulated worlds without the cleanup of the physical world

Availability

Genie 4 is now available to trusted testers as a research preview. Next, it will be offered to Google AI Ultra subscribers (Project Genie) and early partners of the Gemini API / Google AI Studio. You can join the waitlist. The detailed technical report is on the official DeepMind blog.

Limitations

DeepMind did not shy away from the problems. The environment remains coherent within short sessions, but over time visual details and object persistence drift. Complex physical interactions, real-time audio accuracy, text rendering, spatial accuracy, and long-term consistency are all still being researched. All videos and audio generated by Genie 4 will be stamped with a SynthID digital watermark.

Reactions in the Comments

One commenter recalled: "I used to wait five minutes for a ferry to load in The Oregon Trail on a school computer. Now I type a sentence and can walk into a world that talks back." Others were not convinced: "The voice still sounds a bit robotic." Someone asked about consistency: "If I leave an area and come back, will it still be the same?" A game developer remarked: "Last night I spent three hours tweaking a CSS grid. A model that can simulate a room is an obsession from another species."