Skip to content
NewsModel release

Google DeepMind's Genie 4 generates playable worlds with native audio

· by

The short version

Genie 4 holds a consistent, explorable scene for up to two minutes, twice Genie 3's limit, and generates matching music and dialogue as the scene renders. It ships into a world-model market that just saw three acquisitions in two months.

Google DeepMind's Genie 4 generates playable worlds with native audio

Google DeepMind announced Genie 4 on October 27, an interactive world model that generates music, ambient sound, and dialogue in sync with a scene a user is actively exploring and editing by typed prompt, per DeepMind's announcement.

What shipped

Genie 4 holds a consistent scene for up to two minutes, double the limit of Genie 3, and adds improved physics and more humanlike characters, per TechCrunch. A user can type a prompt mid-simulation to change the weather, move the camera, or add a new person or object, and the model adapts the edit to match the existing scene rather than resetting it. DeepMind built Genie 4 on top of its Gemini family: Gemini Omni anchors each simulation's appearance and behavior in real-world references, while other Gemini models handle the reasoning that keeps the scene coherent as it runs. In a research experiment tied to Gemini Robotics, a Gemini 4 Argon model broke a task into steps that Genie 4 then simulated in real time, a possible path to testing robot behavior before it runs on physical hardware. Access starts with a research preview for trusted testers; Google plans to extend it to Google AI Ultra subscribers in Project Genie and to developer partners in the Gemini API and Google AI Studio in the coming weeks.

A crowded field just consolidated around three deals

Genie 4 lands weeks after the world-model category went through a wave of buyouts. AMD agreed to acquire Fei-Fei Li's World Labs for $8.2 billion, Amazon entered talks to buy Decart for roughly $7 billion, and Rivian bought Third Dimension to bring driving-simulation work in-house, per TechCrunch's reporting. Runway shipped a competing model, GWM Worlds 2, just two months earlier. Against that backdrop, Genie 4 is Google's bid to stay ahead using a resource smaller labs cannot match: Gemini itself supplying the reasoning layer on top of the pixel generator. DeepMind research director Tejas Kulkarni described the pairing as Gemini "writing code and text" while Genie handles "ingesting that in real time and simulating reality at different levels of abstraction," per the same TechCrunch report. For teams evaluating robotics-sim or game-prototyping tools, the acquisitions mean fewer independent world-model vendors to pick from going forward, and Google's version is the one tied most directly to a frontier LLM.

DeepMind itself flags where the model still falls short: visual details and object permanence can drift the longer a session runs or the more prompt edits stack up, and the company has not yet solved reliable audio accuracy or legible in-world text, per its own announcement. Trusted testers have so far gravitated toward using Genie 4 as a rapid pre-visualization tool for film and game ideation rather than a finished-content generator.

What to watch next

Watch for the rollout to Google AI Ultra subscribers and Gemini API partners in the coming weeks, and whether any game studio or robotics team publishes results from building on Genie 4 rather than the research-preview demos DeepMind has shown so far.

Sources