Google DeepMind Releases Research Preview of World Model Genie 4

On October 28, 2026, Google DeepMind released the research preview of Genie 4, its next-generation world model. According to the official introduction, the model provides real-time interactive world simulation capabilities, natively generates ambient sound, music, and interactive dialogue, and shows improvements in photorealism and physics simulation. The Verge reported that this update includes more realistic worlds, experiences that can be updated in real time via natural-language prompts, generated audio, and a built-in Gemini 3.8 Flash to assist with simulation. Ars Technica reported that, compared with the previous generation Genie 3, Genie 4 improves physics simulation, audio, and the realism of people; the model uses Gemini 3.8 Flash as a guiding agent responsible for breaking down goals and evaluating every output frame, and users can change the weather, add elements, or move the camera mid-interaction. Regarding session length, Ars Technica said world consistency only lasts a few minutes before the model forgets its initial state, while The Verge said the experience lasts about 2 minutes, double Genie 3's 1 minute. TechCrunch reported that Genie 4 generates the most realistic interactive worlds yet, adding real-time audio matched to the scene. The Decoder reported that Genie 4 generates interactive scenes frame by frame in real time, for the first time generating sound directly and supporting real-time character dialogue and reactions; the model is based on Gemini Omni, with a Gemini model planning the simulation in the background to maintain world consistency—for example, Gemini 3.8 Flash breaks down multi-step goals for execution. The model is currently released as a research preview, with no further news on commercialization or open access plans.

AI overview of reporting · Updated 14 min前

Not enough continuous observations to show a trend yet.