DeepMind's Genie 4 simulates interactive worlds with native audio

Imagine typing a quick scene description and immediately getting an interactive 3D world where you can walk around with your arrow keys, hear the wind through the trees, and actually strike up a live conversation with a stranger on the street.

That is essentially what Google DeepMind introduced on October 27 with Genie 4, its latest world simulation model. While last year's Genie 3 proved that generative models could create playable environments, this fourth generation takes a massive leap by building native, synchronized audio straight into the simulation.

Directing the simulation in real time

Most video and world models remain mute rendering engines. Genie 4 changes that by natively generating contextual soundscapes, from ambient waves and background music to real-time interactive speech from simulated characters.

Users can navigate through environments using standard keyboard controls like WASD or arrow keys. If you want to tweak the setting, you simply type a prompt mid-simulation to change the weather, adjust camera angles, or drop a new object into the scene. The world smoothly adapts without breaking continuity.

Behind the curtain sits Google's broader Gemini ecosystem. Gemini Omni anchors the physical grounding and aerodynamic realism, while reasoning models like Gemini 3.8 Flash act as a real-time director. In one demo, an Apptronik Apollo robot played a game of checkers against a human inside a purely imagined room. The board stayed coherent because a Gemini 4 Argon model was continuously planning each move behind the scenes before feeding the steps to Genie to render.

DeepMind also doubled the model's scene coherence window to roughly two minutes, up from Genie 3.

The cracks in the matrix

Despite the cinematic polish, Genie 4 is still very much a research prototype.

DeepMind openly notes that object permanence and fine visual details begin to drift as simulations run longer or endure multiple complex edits. Spatial accuracy, legible text rendering, and reliable audio fidelity remain works in progress. To mitigate misuse, all outputs carry invisible SynthID watermarks.

For now, the model is restricted to a small cohort of trusted testers across film, gaming, and robotics. Google plans to expand access to Google AI Ultra subscribers inside Project Genie and early access partners using the Gemini API in the coming weeks.

Why it matters

The race around world models has heated up fast. Runway launched its interactive GWM Worlds 2 a couple of months back, AMD moved to acquire Fei-Fei Li's World Labs, Amazon is circling Decart, and Rivian snapped up Third Dimension for autonomous vehicle training. DeepMind had to show something formidable to hold the spotlight.

We think Google's architecture is the real story here. Instead of forcing a single model to handle both high-level causality and pixel-level simulation, they are pioneering a hybrid approach. Gemini handles the narrative logic and task planning, while Genie renders the tactile physical world. That division of labor makes simulated causality far more reliable than purely statistical frame prediction.

For creators and game developers, this is not ready to replace standard game engines, but it could dramatically alter pre-visualization and rapid prototyping. Filmmakers can test complex camera moves inside a living sketch of a set before spending real production budgets.

The catch, of course, is accessibility. Running this requires eye-watering compute, which explains why general availability is tethered to Google's most expensive Ultra tier. Still, Genie 4 makes it clear that generative AI is officially graduating from static clips to living, explorable software.