Google DeepMind Releases World Model Genie 4: Real-Time Interactive Worlds With Native Audio

What happened

DeepMind positions Genie 4 in its official blog as a new generation of world model: after a user sets an event in motion, the model simulates how the world responds next, how it sounds, and how people move and speak. Gemini Omni anchors appearance and behavior in the simulation, while other Gemini models supply the reasoning that keeps the scene coherent; DeepMind says understanding and simulating the world is a crucial aspect of building artificial general intelligence.

Four capabilities make up this update: a running world can be rewritten in real time with natural language, so weather, camera, a new person or a new object can join while the simulation continues, with Gemini adapting the prompt to the surrounding scene and steering the scene between inputs to keep it coherent; on physics and people, the model simulates real-world physics such as aerodynamics and collisions, with people moving more naturally and steerable in real time; on audio, Genie 4 is DeepMind's first real-time world model to generate audio natively, with wind, waves and character dialogue changing in real time with inputs and events; Gemini 3.8 Flash acts as a guiding agent in the loop, observing the scene, breaking goals into steps and reviewing frames for progress, adapting its plan as the scene shifts, and DeepMind also demonstrated Gemini 4 Argon combined with Genie 4 simulating Apptronik's Apollo humanoid playing checkers.

The rollout runs in three steps: the research preview goes to trusted testers first, then Google AI Ultra subscribers through Project Genie, with early access partners of the Gemini API and Google AI Studio also in line, and Project Genie has opened a waitlist. DeepMind says all video and audio generated by Genie 4 carries a SynthID watermark; it also acknowledges the model remains an early experimental model, with visual details and object permanence liable to drift over longer runs, after multiple complex edits or while navigating through a scene, and complex physical interactions, real-time audio accuracy, legible text rendering and spatial accuracy still under research. Reports put scene coherence at about two minutes, double Genie 3.

Key facts

1. Release: Google DeepMind announced Genie 4 in an official blog post on Oct 27, authored by Raia Hadsell, Diego Rivas and Hang Qi.
2. Form: a real-time interactive simulated world, with music, ambient sound and interactive dialogue generated alongside the visuals, not a video clip rendered first and dubbed later.
3. Live rewriting: a running world can be rewritten with natural language prompts changing weather, camera, objects and people, with Gemini adapting prompts to the scene and steering continuously.
4. Guiding loop: Gemini 3.8 Flash observes the scene, breaks goals into steps and reviews frames for progress inside the loop; Gemini Omni anchors appearance and behavior.
5. Access: the research preview goes to trusted testers first, then to Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio, with a waitlist open.
6. Watermark and limits: all generated video and audio carries a SynthID watermark; scene coherence is reported at about two minutes, with details and object permanence liable to drift over longer runs.
7. Uses: the testing directions DeepMind lists include rapid scene generation for film pre-visualization, game prototyping and robot training environments.