Google DeepMind unveils Genie 4 world model with real-time editing and audio

Google DeepMind has announced Genie 4, the next version of its Genie world model. It is rolling out first to a small group of trusted testers as a research preview. The company says the model generates interactive environments in real time. Google highlights four changes. Users can change a running simulation with plain-language prompts. The model simulates physics and characters more convincingly than before. It is the first real-time world model to generate audio natively. Gemini 3.8 Flash guides the simulation as an agent. The Google DeepMind post, written by Raia Hadsell, Diego Rivas and Hang Qi, says access will widen in stages. The first step is Google AI Ultra subscribers in Project Genie, along with early access partners of the Gemini API and Google AI Studio. The company is gathering feedback before a broader release to consumers and developers. Interested users can join a waitlist in Project Genie.

Real-time editing is one of the changes Google emphasizes. According to the post, a user can type a request such as changing the weather, moving the camera or adding a new person or object. Gemini adapts that prompt to fit the existing scene. Between user inputs, Gemini also monitors the scene and steers it so the world keeps progressing coherently. The demonstrations were recorded with standard keyboard controls for navigation and camera movement. Text prompts steered the scene as it evolved.

The architecture shows how Genie 4 draws on Gemini. Genie 4 is built on capabilities from the Gemini family. Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models supply the reasoning needed to keep a scene coherent and continuous. The most specific mechanism is Gemini 3.8 Flash, which the post calls a guiding agent in the loop. As Genie 4 generates a world, Gemini observes the scene, breaks a goal into a sequence of steps, and reviews incoming frames to check that each step succeeded. It revises its plan when the world shifts or a user intervenes. The post says this is what allows worlds with reliable causality and multi-step behavior.

On physics and people, the claims rest on Google's own demonstrations. The post says Genie 4 simulates scenarios such as aerodynamics and collisions. Its most dramatic change, according to the post, is how it models people, producing characters that move naturally and can be steered interactively. The examples shown include a watermelon on a stone plinth, a falling maple seed, a crowded train platform, and a dragon-racing first-person sequence, all presented with the prompts that produced them. The post does not offer benchmark scores, error rates or side-by-side comparisons with earlier Genie versions. The claims of improved realism therefore rest on the company's own demonstrations.

The audio claim comes with a similar caveat. Google writes that Genie 4 is its first real-time world model to generate audio natively. It produces ambient sound, such as wind in trees, along with interactive dialogue that adapts on the fly. The post describes audio-like music, ambient noises and speech being generated alongside the visuals rather than added afterward. It does not say how long the dialogue can run, how many characters can speak at once, or how closely the speech is synchronized with lip movement.

Jack Clark, who writes the Import AI newsletter, pointed out that the current quality is rough. He described the technology as quite janky, pointing to robotic voices and a dreamlike inconsistency in the visualized environments. He framed the broader appeal as generating custom worlds for entertainment and possibly data.