Google DeepMind unveils Genie 4 real-time interactive world model

Google DeepMind has introduced Genie 4, a foundation world model designed for real-time interactive simulation. Built on the Gemini model family, Genie 4 generates dynamic, physics-consistent 3D interactive environments on the fly with real-time natural language steering. Users can modify terrain, atmospheric conditions, physical constants and character behaviour through continuous prompt streams.

Unlike earlier world models that operated purely on visual frame-by-frame dynamics, Genie 4 incorporates native audio generation, ambient acoustics and multi-agent spoken dialogue directly into the simulation stream. The system synthesises reactive sound effects and spatial acoustics that match physical events inside the simulated space without relying on external text-to-speech or audio synthesis pipelines.

Steering operates through keyboard, gamepad or continuous spoken instructions, allowing a participant to direct autonomous agents and environment dynamics simultaneously. The architecture runs at interactive frame rates by combining compressed spatiotemporal latent dynamics with fast autoregressive decoding, maintaining visual and physical consistency across extended multi-minute interaction loops.

Google is rolling out Genie 4 as a research preview to trusted testers. The company confirmed that access will expand to Google AI Ultra subscribers within Project Genie, with programmatic access opening to early access partners through the Gemini API and Google AI Studio via an application waitlist.