Google DeepMind releases Genie 4, a world model with native audio and text steering

Google DeepMind has announced Genie 4, the next version of its real-time world model. A world model produces an interactive environment frame by frame, and a user can move through it while it is being generated. Genie 4 adds three things: sound is generated together with the picture, physics and human movement look more realistic, and plain-text prompts can change the scene while the simulation is running.

The model is not widely available. Google DeepMind says Genie 4 is a research preview for trusted testers. The company plans to bring it first to Google AI Ultra subscribers in Project Genie, and then to early-access partners in the Gemini API and Google AI Studio. It has not given a general release date. The announcement includes no benchmarks, resolution, frame rate or latency figures, so the claims below are the company's own descriptions and have not been measured independently.

What changed from Genie 3

Genie 3 was announced in August 2025. Google DeepMind calls Genie 4 "our first real-time world model to generate audio natively." Ambient sound such as wind and waves is generated together with the visuals, and so are music and dialogue. The company says characters can hold conversations that adapt as they go: "Characters converse and react to your inputs or events in the world, all in real-time."

On physics, the post says Genie 4 "simulates real-world physics scenarios, such as aerodynamics, collisions, and more." The demonstration prompts include a watermelon, a falling maple seed and a train station. The company also says characters "move naturally, respond organically to their environments" and can be steered while the simulation runs.

Genie 4 draws on the Gemini family of models. According to Google DeepMind, Gemini Omni ties the visuals to how things look and behave in the real world, and other Gemini models handle the reasoning that keeps scenes coherent.

Steering a running world with text

The biggest change to how people use it is mid-simulation control. You move with the keyboard as before, but you can also type an instruction while the world is running. Examples include changing the weather, moving the camera, or adding a person or an object. Google DeepMind says Gemini adapts each prompt to the surrounding scene and keeps the world moving forward between inputs.

The company describes a second mode in which Gemini 3.8 Flash acts as "a guiding agent in the loop." In this mode Gemini watches the frames, breaks a goal into a sequence of steps, checks each step, and plans again when the world changes or the user steps in. One model generates the world while another model directs it toward a goal.

Google DeepMind presents this as an early form of what it calls hybrid world models. Gemini would "map the structure and plan the narrative timeline," and Genie would render how the world looks, sounds and responds at the pixel level.

Robot training and testing

The post says Genie 4 can help test how AI systems navigate environments before they are deployed in the real world. Google DeepMind says it has experimented with combining Genie 4 with Gemini 4 Argon and with Gemini Robotics. In one example, the two models "work together to simulate Apptronik's Apollo humanoid playing a game of checkers." Argon splits the task into steps and passes them to Genie as text captions, and Genie renders each step.

In this setup a language model writes the plan and the world model shows what carrying it out would look like. The robot does not have to touch a real board. The post does not say whether policies trained or tested this way perform better on physical hardware.

Early users and stated limits

Google DeepMind says its trusted testers work in creative industries, journalism, education, gaming and engineering. It lists several uses:

• Film and creative direction: quick previsualisation, and generating several "takes" of the same scene.
• Photojournalism: exploring how "a single 2D archival photograph can become a grounded simulation."
• Simulating events such as floods or wildfires.
• Game and experience design: going from an idea to an explorable prototype through prompts.
• Enterprise scenario generation: partners in several industries are evaluating it.

The company is direct about the limits. The model keeps environments coherent over short sessions, but "visual details and object permanence can still shift or drift over time." This happens most over longer runs, after many prompt edits, or while the user is moving around. It lists the areas that are still open: complex physical interactions, accuracy of real-time audio, readable text, spatial accuracy and long-term consistency. All generated video and audio carries a SynthID watermark. Google DeepMind says the model went through internal safety evaluations before the preview.

What it means for work

In game design, Genie 4 targets the early prototyping stage: a designer can describe a level and walk through it with sound before anyone builds assets or writes code. Film previsualisation and storyboarding could move the same way, with directors generating rough takes of a scene instead of commissioning them. For robotics teams, this is another way to rehearse tasks in simulation, with the plan written in language rather than scripted by hand. That could reduce some of the effort that now goes into building simulation environments. Two things remain open. Drift over long sessions limits how far these worlds can stand in for a production engine or a physics simulator that must stay consistent. And without published metrics, nobody outside Google can yet judge how closely the simulated physics matches reality. For now the model looks most useful for fast, rough exploration, while precise simulation still needs the tools people use today.