Google DeepMind's blog, dated October 27, 2026, announces the world model Genie 4. The byline names Raia Hadsell, Diego Rivas, and Hang Qi. The post describes Genie 4 as real-time interactive world simulation, and it says picture and sound are generated together. At the start, the sound is described as music-like audio, ambient noise, and interactive dialogue. That is the company's own release note, not a measurement from outside.
[1]
The model is tied to the Gemini family. The post says Gemini Omni anchors the simulations in real-world appearance and behavior, while other Gemini models supply the reasoning used to keep a simulation coherent and continuous. A person can set an event in motion, and the page says Genie 4 then simulates what happens next: how the world responds, how it sounds, and how the people in it move, react, and speak. During a run, natural language can change the weather, move the camera, or bring in a new person or object. The post says Gemini adapts that prompt to the surrounding scene, and between inputs it watches the scene so the world keeps moving forward. The recordings on the page use standard keyboard controls for navigation and for the camera, plus text prompts to steer. The physics passage names aerodynamics and collisions. The post treats the simulation of people as the sharpest change in this version: people can move, respond to the environment, and be steered while the simulation is running.
[1]Audio is the part the post marks as a first for this line of models. The text says Genie 4 is the company's first real-time world model to generate audio natively, and that every world carries the sound of its own setting. Characters can talk, and they can react in real time to inputs or to events in the scene. A later section names Gemini 3.8 Flash as a guiding agent in the loop. While Genie 4 generates a world, that Gemini model watches the frames, breaks a goal into steps, and checks the incoming frames to see whether each step succeeded. If the scene changes, or if a person steps in, the plan changes with it. The post also describes an experiment that places Gemini 4 Argon with Genie 4 and with Gemini Robotics. In the example, Argon splits a task into steps for Genie to render, and those steps are inserted as text captions while the simulation runs. The caption says Gemini 4 Argon and Genie 4 together simulate Apptronik's Apollo humanoid playing a game of checkers.
[1]The rollout is narrower than the list of abilities. The post says Genie 4 has already begun to go out as a research preview for trusted testers, so feedback can be collected before a release aimed at consumers and developers. The later group named on the page is Google AI Ultra subscribers in Project Genie, plus early-access partners of the Gemini API and Google AI Studio. The page has a waitlist. It does not say on which day Ultra subscribers or API partners get access. Testers are described as cohorts in the creative industries, journalism, education, gaming, and engineering. The uses are also written as work in progress: film people changing camera, light, dialogue, and action inside a live simulation; photojournalists trying to turn one archival photograph into an environment a person can enter; game developers trying style and interaction in language before deciding whether to write engine code. The page does not say how many finished pieces those tests produced.
[1]The limits sit in the post's own section, and they keep this announcement inside a preview. Genie 4 is described as an early experimental model. Over a short stretch, an environment can stay coherent, but visual detail and whether an object is still where it was can drift, especially after a simulation runs longer, after several complex edits to the prompt, or while moving through the scene. Complex physical interactions, whether real-time audio is accurate, whether readable text can be generated, whether space stays accurate, and whether a long session stays consistent are all listed as research that is still underway. On safety, the post says internal safety evaluations were done before the preview, in order to block harmful or malicious content, and that the model was given first to trusted tester partners. Every video and every audio clip generated by Genie 4 will carry SynthID. The post calls that an imperceptible watermark for identifying generated material. The page does not publish the bar those evaluations had to clear, and it does not say how long a typical simulation stays steady.
[1]