Google DeepMind has unveiled Genie 4, an AI model that offers a real‑time interactive simulation of a real world, with native sound and increased realism. The rollout began as a research preview with trusted testers.
The system lets users generate dynamic environments and move through them, piloting scenes on the fly. Meanwhile, the model simultaneously computes physics, character behavior, and soundscapes. Sundar Pichai, CEO of Alphabet and Google, called Genie 4 “our most realistic and interactive world model.”
Google DeepMind, an Anglo‑American research lab specializing in artificial intelligence, is a subsidiary of Alphabet Inc. Founded in the United Kingdom in 2010, it was acquired by Google in 2014, then merged with Google AI’s Google Brain division to become Google DeepMind in April 2023. Its headquarters are in London, and the company has research centres in the United States, Canada, France, Germany and Switzerland.
Real‑time piloting and sound generated alongside the image
According to Google, users can explore simulated environments using standard keyboard controls and alter the world with simple plain‑text prompts. They can change the weather, adjust camera angles, or introduce new characters and objects. Google also says Genie 4 represents a significant leap in physical accuracy, with complex interactions such as aerodynamics, collisions, or environmental resistance, and that digital characters move and react instantly to commands.
Google presents Genie 4 as the first real‑time world model capable of natively generating synchronized sound as images are rendered. It produces contextual ambiances, such as wind rustling through trees or waves crashing, as well as background music. It also handles interactive spoken dialogue: characters converse and respond to user requests as well as to environmental events.
A model built on the Gemini family
Genie 4 is not a standalone graphics engine. It relies on the multimodal reasoning capabilities of the Gemini family. Gemini Omni grounds each simulation in appearance and physical dynamics drawn from the real world, while Gemini reasoning models preserve scene logic so that events maintain cause‑and‑effect coherence. The Gemini 3.8 Flash model acts as a piloting agent: it follows incoming images, breaks goals into steps, verifies that each action succeeds, and adapts the plan whenever a change occurs in the scene or a user intervenes.
Google notes that this combination enables modeling of multi‑step complex events and testing AI system behavior in different environments before deploying them in the real world. The company also reminds that “understanding and simulating the world is a crucial aspect of building general AI (AGI).”
Early feedback still experimental
Genie 4 remains, according to Google, an experimental model at an early stage. The rollout initially targets Google AI Ultra subscribers as part of the Genie project, as well as partners with early access to the Gemini API and Google AI Studio. Google has worked with testers from creative fields, journalism, education, video gaming and engineering. Filmmakers, for example, can capture multiple takes of the same scene in real time, while photojournalists explore how a 2D archival photograph can become a realistic simulation of its environment. Other teams are examining visualization of short‑duration environmental events, such as a sudden flood or a forest fire.
The model also fits within the work of Alphabet’s autonomous‑driving subsidiary. In February 2026, Waymo, owned by Alphabet, unveiled Waymo World, a model that leveraged Genie 3’s capabilities to build virtual environments from text prompts. The system generated synthetic driving images as well as depth‑perception data comparable to that from cameras and lidar sensors mounted on vehicles.
