Genie 4 generates interactive worlds, now with native audio

Google DeepMind released Genie 4, its new world model that builds interactive, playable environments in real time. The headline change is audio. DeepMind says Genie 4 is the first real-time world model to generate sound natively, including music, ambient noise, and interactive character dialogue, rather than adding it as a separate layer. You can steer a running simulation with plain language, changing the weather, moving the camera, or dropping in new people and objects, and the system tries to keep the world coherent across those edits.

What ties it together is Gemini acting as a guiding agent. A Gemini 3.8 Flash model watches the scene, breaks a goal into steps, checks whether each step worked, and rewrites its plan when the world shifts or a user steps in. DeepMind is also testing this with robotics: one demo pairs Genie 4 with Gemini Robotics to simulate Apptronik's Apollo humanoid playing checkers. The simulation layer, Gemini Omni, anchors appearance and behavior to the real world.

DeepMind is honest about the limits. Visual detail and object permanence still drift over longer sessions or after many edits, and the team lists physics, legible text, and long-term consistency as open problems. There are no benchmarks in the announcement. Genie 4 ships as a research preview for trusted testers, with a waitlist and gradual access through Project Genie and the Gemini API. All output carries SynthID watermarks.

Why it matters

If you work on robotics or agent training, a controllable world model with native audio is a cheaper way to generate and test scenarios than collecting real-world data. The drift problems mean you should still treat long simulations with caution before relying on them.