Google DeepMind releases Genie 4 world model with audio, speech and two-minute consistency

What Google released

Google DeepMind released Genie 4 on Tuesday, 27 October, in a research preview. Genie is a world model, a type of AI that simulates an environment users can walk through with a keyboard and change with written instructions. Instead of designing every element in advance, the system generates each frame in real time from a text description. The first release is aimed at 'trusted testers', the people who receive early access to Google products. Google presents the technology as a step toward artificial general intelligence, able to handle very diverse tasks beyond video games.

Genie's development path

Genie 1 generated controllable two-dimensional environments and was presented in February 2024. Genie 2 moved to 3D in December of that year, and Genie 3, announced in August 2025, allowed more realistic exploration in real time at 720p and 24 frames per second.

None of these models has been widely available. Genie 3 could be accessed through Project Genie, an experimental web app that combined the model with Nano Banana Pro and Gemini. Users had to be at least 18 years old and hold a Google AI Ultra subscription, and explorations were limited to 60 seconds.

What the new version adds

Compared with Genie 3, DeepMind says Genie 4 improves the simulation of collisions and aerodynamics, as well as movement and human reactions. The most prominent addition is native audio, generated together with the images and built into the environment, such as wind through trees or waves breaking on a beach. The model also produces human speech and interactive dialogue, so characters can talk to each other and react to events in the simulation. Users can type prompts mid-simulation to change the weather or camera angle, or to add new people and objects.

The consistency window is the clearest numerical change. Genie 4 keeps a scene consistent for up to two minutes, double the limit of its predecessor, which allowed about one minute.

Genie 31

Genie 42

Genie 3

1 minutes

Genie 4

2 minutes

Data from the article text

Gemini's role and known limits

Gemini is more closely integrated with Genie 4. The system uses Gemini Omni capabilities to improve realism and relies on Gemini 3.8 Flash as a supervisor that observes the scene, plans steps and checks frames while Genie creates the visuals. In a research experiment tied to Gemini Robotics, the system broke tasks into steps that Genie 4 simulated in real time, a possible way to test robot scenarios before they happen in the physical world. Tejas Kulkarni, research director at Google, described the direction during a briefing.

“What we'll increasingly see in the future are these hybrid world models, where you have a combination of Gemini writing code and text”

The limits are clear. Google warns that as simulations run longer, details and object permanence can shift over time. Reports on the launch note that the model begins to forget what it has shown after about two minutes.

Uses, rivals and access

Google pitches Genie 4 for pre-visualization for filmmakers, prototyping game ideas and building training scenarios for robots and AI agents, which can learn the consequences of their actions before acting in the physical world. Trusted testers are pointing to prototyping films as a possible new use.

Bloomberg frames the release against investor concern, noting that the technology has rattled investors in video game companies. The field is consolidating: AMD is acquiring World Labs, the company led by Fei-Fei Li, Amazon is in talks to buy Decart, and Rivian acquired Third Dimension to bring simulation capabilities in-house for self-driving testing. Runway had already released GWM Worlds 2, a playable model that includes audio generation.

Genie 4 will be 'coming soon' to Google AI Ultra subscribers, and Google has opened a waitlist where people can apply to try it through Project Genie.