Google DeepMind unveils Genie 4, a real-time world model with native audio, text steering and two-minute scenes

Google DeepMind released Genie 4 as a research preview on Oct 27, 2026, a little over a year after Genie 3. The blog post (Raia Hadsell, Diego Rivas, Hang Qi) describes a world model that renders and sounds out an interactive scene in real time, lets users change the running world with natural-language prompts, and uses Gemini models as a “director” that plans multi-step events and checks the generated frames. DeepMind pitches it for gaming, interactive learning and robotics; a demo shows an Apptronik Apollo humanoid playing checkers inside a simulation, and a Gemini Robotics experiment had Gemini 4 Argon break a task into steps that Genie 4 then simulated.

TechCrunch reports scenes stay consistent for up to two minutes, double Genie 3’s limit, and quotes Google research director Tejas Kulkarni: “What we’ll increasingly see in the future are these hybrid world models”. Trusted testers mentioned film prototyping and storyboarding as uses. Resolution, frame rate and pricing were not published.

Genie 4 joins sound to playable world generation in one real-time model and ties it to Gemini’s reasoning, the “hybrid” design DeepMind says world models are moving to. Competitors are consolidating (AMD buying World Labs, Amazon reportedly in talks for Decart, Runway’s GWM Worlds 2 with audio), and world models are the main simulator pitch for training robots and agents. Access is still limited to testers, so outside evaluations do not exist yet.