Google DeepMind has announced Genie 4 as a research preview for trusted testers. The company says the model generates explorable worlds from text and image prompts in real time, with visuals, ambient sound, music and character dialogue produced together.

Users can navigate with keyboard controls and steer events through plain-language prompts, such as changing weather, moving the camera or adding an object. According to DeepMind, Genie 4 simulates elements including collisions, aerodynamics and human movement, while characters can respond to their surroundings and hold conversations.

A central design change is the use of a Gemini model inside the simulation loop. DeepMind says Gemini 3.8 Flash interprets requests, breaks goals into steps, checks incoming frames and revises plans when users intervene or the scene changes. The company describes this as an early example of a hybrid world-model design in which a language model handles structure and planning while Genie generates pixels and sound.

DeepMind has not disclosed Genie 4’s parameters, resolution, frame rate or benchmark results. The company says sessions last a few minutes, with scene consistency reported for up to two minutes in a TechCrunch briefing. It also lists visual drift, imperfect object permanence, limited real-world accuracy and unresolved audio and multi-agent interaction issues among the model’s limitations.

Access currently begins with trusted testers. DeepMind says availability is expected next for Google AI Ultra subscribers through Project Genie and for selected Gemini API and Google AI Studio partners. Genie 4 outputs will carry SynthID watermarks, according to the company.