Google DeepMind Unveils Genie 4 World Model With Native Audio and Real-Time Editing
Smartor 편집팀
October 27, 2026
Featured image: AI-generated explanatory illustration of world simulation, not an actual Genie 4 interface or output.
Google DeepMind announced Genie 4 on its official blog on October 27 and said it had begun rolling out the real-time world model as a research preview for selected trusted testers. The model generates visuals and sound together, allowing people to alter an unfolding virtual scene with text prompts while producing ambient audio, music and dialogue from simulated characters.
World models generate the next scenes and changes in a virtual environment in response to user input. DeepMind described Genie 4 as its first real-time world model with native audio generation. Its examples include environmental sounds such as wind and waves, alongside characters that speak in response to user inputs or events within the simulation.
According to the company, natural-language prompts can change the weather, move the camera or introduce people and objects while a simulation is running. DeepMind also reported improvements in physical interactions, including collisions, and in the depiction of human movement. These are the company’s descriptions of its capabilities, not a claim that it reproduces reality accurately in every situation.
Gemini observes the scene and adjusts its progression
Genie 4 works with models from Google’s Gemini family. DeepMind said Gemini Omni helps ground the appearance and behavior of a simulation, while Gemini 3.8 Flash observes the generated scene, breaks goals into steps and checks whether those steps have succeeded. It adjusts its plan when the environment changes or a user intervenes.
The company also showed a research experiment combining Gemini 4 Argon and Genie 4 to simulate a humanoid robot playing checkers. Trusted testers have been exploring film previsualization and storyboarding, as well as early prototypes for games and interactive experiences, DeepMind said. Those examples describe research and testing, rather than the launch of a finished commercial production tool.
Limited preview, with consistency still a challenge
DeepMind characterized the model as early and experimental. Visual details and the ability to keep objects consistent can drift during longer simulations or after repeated complex edits. It listed complex physical interactions, audio accuracy, legible text, spatial accuracy and long-term consistency as continuing research challenges. TechCrunch’s October 27 report also said the team acknowledged these limitations during a briefing.
Access is currently limited to trusted testers. DeepMind said availability would expand to Google AI Ultra subscribers through Project Genie and to early access partners in the Gemini API and Google AI Studio. The official announcement did not give a specific date for a general release. The company said generated video and audio would carry SynthID, its imperceptible watermarking technology for identifying AI-generated content.