Google DeepMind has released Genie 4, a real-time interactive world model. While a simulation is running, users can use natural language to change the weather, move the camera, and add people and objects to the scene, and the picture keeps playing forward without being regenerated. DeepMind calls it the first real-time world model to generate audio natively: ambient sound, music and character dialogue are produced in sync with the visuals, and characters can respond in real time to user input or to events in the scene.
"Genie 4 is our first real-time world model to generate audio natively."
What it does, and who gets it
According to the official description, the Genie 4 upgrade centers on three areas. The first is physics, with support for scenarios such as aerodynamics and collisions. The second is characters, whose movement is more natural, who react to their surroundings, and who can be controlled by the user in real time. The third is multi-step events: Gemini 3.8 Flash acts as a guide agent, breaking the user's goal into steps, checking frame by frame whether each step has been completed, and revising the plan when the user steps in mid-way or the scene changes. DeepMind also says it is experimenting with connecting Gemini 4 Argon to Genie 4 for robotics-related simulation.
Availability comes in three stages. Right now it is a research preview for trusted testers. Next come Google AI Ultra subscribers (through an entry point called Project Genie) and early partners on the Gemini API and Google AI Studio; the page carries a waitlist form. Generated video and audio will both carry SynthID watermarks.
The shortcomings DeepMind lists itself
The announcement contains no quantitative metrics. Resolution, frame rate, playable session length and input latency are all missing, and there is no numerical comparison with the previous generation, only qualitative phrases such as improved realism.
DeepMind does spell out the limitations. After a session grows long, a complex prompt has been revised several times, or the user has wandered far across a scene, visual detail and object permanence "still shift or drift over time." Complex physical interactions, the accuracy of real-time audio, legible text rendering, spatial precision and long-term consistency are all listed as open research problems.
How readers in China should read this
The first access channels for Genie 4 are the AI Ultra subscription and early Gemini API partnerships. Neither is available in mainland China, so developers there can for now only watch the demo videos.
The domestic field is not empty. Matrix-Game 4.0, which Kunlun Tech open-sourced in mid-October, is pitched at 60 frames per second on a single RTX 5090, and its weights can be downloaded and run locally. Other teams have also run world models at around 50 frames per second on domestic NPUs. These projects publish comparable metrics such as frame rate, whereas Genie 4 this time does not.
The gap is more likely to lie in two places. One is native audio and character dialogue: most open-source efforts in China currently output only visuals. The other is using Gemini as a planning agent that checks a multi-step task frame by frame, which requires a strong multimodal model watching alongside and is hard for teams that only build video generation to replicate. Only once Genie 4 reaches the API stage and publishes latency and duration figures will there be a table on which the two sides can be read together.
Sources: Google DeepMind official blog, CocoLoop; the availability scope, the model used by the guide agent and the known limitations follow DeepMind's official statements, and frame rates for domestic world models are based on each project's public data.