Google DeepMind has introduced Genie 4, a world model built to generate environmental sound, music and interactive dialogue natively while a scene is unfolding. The system also lets users change the setting in real time with natural-language prompts. Outside reports said coherent interaction lasts about 2 minutes, roughly double the approximately 1 minute reported for Genie 3.
In the early hours of Oct. 28 Taipei time, Google DeepMind posted a roughly 147-second demo on X and described Genie 4 as its most realistic and most interactive world model so far. The company is not opening it broadly to the public yet. In a blog post released at the same time, DeepMind said the model is being offered first as a research preview to trusted testers.
What Genie 4 adds
DeepMind said Genie 4 is natively multimodal. As visuals progress, the model can generate music, ambient sound and interactive dialogue at the same time. The company said this is the first time audio has been generated natively inside a real-time world model rather than added later as a separate track.
On the navigation side, Genie 4 is tied to the Gemini family. DeepMind said Gemini Omni helps anchor simulations to real-world appearance and behavior, while other Gemini models supply coherent reasoning. Users can change weather, camera views and add people or objects through natural-language prompts on the fly. During pauses and transitions, Gemini also monitors the scene and corrects its direction.
DeepMind also highlighted upgrades to physics and character behavior. Aerodynamics and collisions are described as more realistic, while character movement is more natural and can be guided in real time. In the company demo, movement and camera control are handled with a keyboard, and text prompts are used to push the scene forward. DeepMind said the underlying process is pixel-level real-time generation rather than playback of prebuilt game assets.
Why Gemini and robotics are part of the setup
DeepMind said it is pairing Genie with Gemini to simulate complex, multi-step events. In practice, Gemini 3.8 Flash acts as a guiding agent in the loop, watching the scene, breaking down goals and checking whether each step succeeds. If the environment changes or a person intervenes, the plan is revised on the spot.
The company also referred to experiments involving Gemini Robotics and Gemini 4 Argon. According to DeepMind, Argon breaks a task into steps and then uses live text captions to direct Genie 4 as it renders the scene. One example given was simulating a humanoid robot playing chess. For people focused on on-chain AI and virtual-world economies, the report framed this more as a testable sandbox becoming mature before questions of content and assets are addressed.
Who is testing it, how long it lasts, and where it still falls short
DeepMind said trusted testers already include people in film and television, news, education, gaming and engineering. Some are using it for previs drafts. Others are turning a historical photo into a space that can be entered and explored. On the gaming side, testers are trying natural-language prompts to explore appearance and interaction before writing engine code.
At the same time, the company stressed that Genie 4 remains an early research effort. It has not locked in what a full end-to-end product will look like.
On interaction length, both TechCrunch and The Verge reported that Genie 4 can stay interactive for about 2 minutes, about twice the duration cited for Genie 3. DeepMind used more restrained wording in its own materials, saying only that the model can remain coherent over short spans, while details and object permanence can drift during longer runs.
DeepMind listed its current limitations directly: complex physics, audio accuracy, readable text, spatial accuracy and long-term consistency are all still under study. In the company’s framing, Genie 4 today looks more like a real-time draft system that can run for several minutes than a stable world that can be explored without limit.
Release plan
DeepMind said trusted testers will get access first. After that, Genie 4 is planned for Google AI Ultra subscribers through Project Genie, as well as early partners using the Gemini API and Google AI Studio. A waitlist is also being opened.
The company added that generated content will carry SynthID watermarking. Safety evaluations and guardrails will follow official policy, with feedback gathered in a smaller circle before access expands.