Google DeepMind's Genie 4 adds native audio and real-time steering to world simulation

Real-time worlds that look and sound real

Google DeepMind has released Genie 4, its latest world model, in research preview. The announcement marks two steps forward for the company's real-time simulation work: audio is now generated natively alongside the visuals — ambient sound, music and dialogue that reacts to what happens — and the worlds themselves are the most photorealistic the company has produced yet.

Users move through Genie 4 environments with standard keyboard and camera controls, and steer what happens with text prompts while the simulation runs. According to DeepMind, a prompt can change the weather, reposition the camera, or drop a new person or object into a scene, and the system reinterprets the instruction so it fits the surrounding context. Between inputs, a Gemini model watches the scene and nudges it forward to keep it coherent.

The model is built on a physical understanding of the world, DeepMind says, and can simulate scenarios such as aerodynamics and collisions. Characters move naturally, react to their environment, and can be steered interactively, and they talk back: dialogue adapts on the fly to user input or to events unfolding in the world. Every environment also carries its own soundscape, from wind in the trees to waves on a beach.

According to TechCrunch, Genie 4 can hold a consistent scene for up to two minutes — twice the limit of its predecessor Genie 3, which arrived as a research preview a little over a year ago.

Gemini as the director

Genie 4 is tightly coupled with Google's Gemini models. DeepMind says Gemini Omni grounds the simulations in real-world appearance and behavior, while Gemini 3.8 Flash acts as a guiding agent in the loop: it observes the generated world, breaks goals into steps, checks incoming frames to confirm each step succeeded, and revises its plan when the user intervenes.

In a research experiment tied to Gemini Robotics, a Gemini 4 Argon model decomposed a user-instructed task into steps that were fed to Genie 4 as text captions while the simulation rendered — demonstrated with Apptronik's Apollo humanoid playing a game of checkers. DeepMind frames this as a way to rehearse robot scenarios in simulation before running them on real hardware.

The company's stated direction is hybrid world models that pair the structured reasoning of code and language models with pixel-level simulation. Tejas Kulkarni, a research director at Google, told a briefing that the future lies in Gemini writing code and text while Genie ingests it in real time, TechCrunch reports.

What testers are doing with it

Early access cohorts span creative industries, journalism, education, gaming and engineering. DeepMind says filmmakers are using it as a live pre-visualization tool, stepping into a scene to run multiple takes and adjusting camera angles, lighting and action on the fly; photojournalists are turning a single archival photograph into an explorable simulation, including models of short-term events like flash floods or wildfires; and game developers are prototyping visual styles and interactive mechanics through natural language before writing engine code.

DeepMind engineer Diego Rivas told reporters that the real-time aspect matters most where maintaining a state of creative flow is useful, such as ideation and storyboarding, according to TechCrunch.

Rollout, limits and competition

Genie 4 is available now to trusted testers, with a rollout in the coming weeks to Google AI Ultra subscribers via Project Genie and to early access partners of the Gemini API and Google AI Studio. All generated video and audio carries DeepMind's SynthID watermark, and the model went through internal safety evaluations before launch.

The team is candid about gaps. Per TechCrunch, details and objects can drift the longer a simulation runs, and the researchers have not yet solved how to guarantee audio accuracy or legible in-world text.

The release also lands in a consolidating field. TechCrunch notes that Runway shipped its second world model, GWM Worlds 2, with audio generation a couple of months ago, while AMD is acquiring Fei-Fei Li's World Labs, Amazon is in talks to buy Decart, and Rivian has bought Third Dimension to bring self-driving simulation in-house.

Why it matters

DeepMind argues that understanding and simulating the world is a key part of building AGI, and Genie 4 points at the practical near-term uses: training environments for robots and agents, rapid prototyping for games and film, and visualizing scenarios that are too costly or dangerous to stage in reality. Natively generated audio combined with real-time steering narrows the gap between generated worlds and authored interactive experiences, and the Gemini-in-the-loop architecture suggests a template — language models planning, world models rendering — that competitors are likely to emulate. For developers, the upcoming Project Genie and API access offers a first chance to test whether live, editable simulation becomes a build target rather than a research demo.