Google previews Genie 4 world model with audio, natural-language editing and robotics tests

Genie 4 is the latest in a series of world models Google launched in 2024. Compared with its predecessor, it is better at rendering fluid movements and other complex details, and it can generate accompanying audio such as the sound of crashing waves.

“The most dramatic leap is how Genie 4 simulates people, creating characters that move naturally, respond organically to their environments, and can be interactively steered in real-time,” Google DeepMind staffers Raia Hadsell, Diego Rivas and Hang Qi wrote in a blog post.

Users can create a virtual world by giving Genie 4 a natural-language description of what they want to render. They can then navigate the simulation and make edits by entering follow-up prompts, such as asking the model to change the weather or add more objects. The model can make such edits without reducing the virtual world’s photorealism or consistency. The natural-language interface is available alongside keyboard controls that can change the position and angle of the camera.

Under the hood, Genie 4 uses several other Google models to improve output quality. It relies on Gemini Omni, a video generator Google introduced in May, to ensure that the visual assets it generates are realistic. Gemini 3.8 Flash, a reasoning model optimized for cost-efficiency, makes individual frames of a simulation consistent with one another.

“Gemini continuously observes the scene, breaks down goals into a sequence of steps, and reviews incoming frames to track the progress and verify that each step succeeded,” the DeepMind staffers wrote. “As the world shifts or the user intervenes, Gemini adapts its plan in the moment.”

DeepMind is studying how Genie 4 can be applied to robotics, combining the algorithm with Gemini 4 Argon, Google’s flagship reasoning-optimized large language model. The project aims to let engineers simulate robots using high-level instructions. In one example, a user could ask Argon to generate synthetic footage of a robotic arm soldering together two car parts. Argon can turn that brief request into dozens of highly detailed, multi-sentence Genie 4 prompts, each describing a different step of the soldering workflow.

Increasing the number and precision of prompts provided to Genie 4 improves rendering quality. That quality boost makes the model’s robot simulations more useful for AI training, where developers train models to control robots by providing visual examples of how to complete tasks.

Google is testing Genie 4 with a limited number of early adopters. It plans to make the model available through its Google AI Ultra subscription for consumers, the Gemini API and Google AI Studio.