Google DeepMind on Tuesday announced Genie 4, its new world model for real-time interactive simulation — and this time, the simulated worlds make sound. The model generates music, ambient noise, and spoken dialogue at the same time as it generates the world itself, according to DeepMind, making it the lab’s first real-time world model with native audio.
A world model is not a video generator. Instead of rendering a fixed clip, it builds a place you can move through, predicting what happens next as you act. With Genie 4, a user types a prompt — change the weather, move the camera, drop a new person into the scene — and the simulation adjusts while it is running. DeepMind says Gemini models adapt each prompt to fit the surrounding scene and keep steering the world between inputs so it stays coherent.
DeepMind lists four advances over its earlier work. The first is that live steering: natural-language edits applied mid-simulation, with Gemini rewriting the request to match what is already on screen. The second is realism in physics and people. Genie 4 simulates aerodynamics and collisions, DeepMind says, and its characters now move naturally, react to their surroundings, and can be steered in real time — the most dramatic leap in the release, in the lab’s own assessment.
The third is sound. Every world carries the ambient noise of its setting, from wind in trees to waves at a beach, DeepMind says, and characters can now hold interactive dialogue that adapts on the fly to prompts and events in the world. The fourth is reasoning: Gemini models supply the logic that lets Genie 4 stage complex, multi-step events rather than single moments.
Gemini steers from inside the loop
The steering runs on two parts of the Gemini family, according to DeepMind. Gemini Omni anchors simulations in how the real world looks and behaves, while other Gemini models provide the reasoning that keeps a scene consistent. Gemini 3.8 Flash acts as a guiding agent inside the loop: it watches the scene, breaks a goal into a sequence of steps, checks each frame as it arrives, and revises its plan when the world shifts or the user intervenes.
DeepMind is also testing the combination with its robotics work. In one experiment described in the announcement, Gemini 4 Argon and Genie 4 together simulate Apptronik’s Apollo humanoid playing checkers, breaking the task into real-time steps as the simulation renders. It is an early research demonstration, not a product — but it shows why DeepMind treats world models as training grounds. The lab says Genie 4 can help build training environments for robots and agents, a use case that echoes recent work elsewhere, including a robot hand that can feel its way to new capabilities.
Trusted testers are already trying the model in other fields, DeepMind says. Filmmakers are using it as a live “draft mode” for pre-visualization, running multiple takes of a scene while adjusting camera angles, lighting, dialogue, and action. Photojournalists are testing whether a single 2D archival photograph can become a grounded simulation of an environment in seconds. Game developers are going from an idea to an explorable prototype without writing engine code first, and enterprise partners are testing rapid scenario generation.
A preview, not a product — yet
For now, Genie 4 is a research preview for trusted testers. DeepMind says it will come next to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio, with a waitlist open for Project Genie.
Scenes hold together for about two minutes, according to reports on the announcement — roughly double the one-minute limit of Genie 3. DeepMind itself does not state a runtime figure in its announcement, so that number should be read as reported, not as the lab’s own specification.
DeepMind is also upfront about what Genie 4 cannot yet do. Visual details and object permanence can still shift or drift over time, especially during longer runs, after multiple complex prompt edits, or while navigating through a scene. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy, and long-term consistency remain active research problems, the lab says. All video and audio the model generates will carry SynthID, Google’s imperceptible watermark, and the preview is gated to trusted testers after internal safety evaluations.
The lab frames world simulation as a step toward artificial general intelligence, calling it “a crucial aspect of building AGI.” That framing sits alongside the limits it lists in the same post: Genie 4 is a striking demo of a simulated world that talks back, but the worlds drift, the text still smudges, and the public cannot try it yet. The more immediate story is narrower and more concrete — AI agents that act in the world rather than just answering questions are gaining a training ground that looks, sounds, and behaves a little more like the real one. As with other frontier models, the question now is how fast the research preview turns into something ordinary users can open.
Sources: Google DeepMind, Google Blog.
