Google DeepMind releases Genie 4, a real-time interactive world model with native audio

Google DeepMind announced Genie 4 on October 27, 2026, as a new world model offering unprecedented realism and real-time interactive simulation.
The model is natively multimodal, generating audio (music, ambient sounds, dialogue) simultaneously with visual simulations, marking the first real-time world model with native audio generation.
Genie 4 integrates with the Gemini family, where Gemini Omni anchors realistic appearance and action, while other Gemini models handle complex reasoning for simulation coherence and continuity.
The system supports real-time steering via natural language prompts, allowing users to change weather, move cameras, or introduce new objects while maintaining scene consistency.

Technical Capabilities

Advanced physical understanding simulates aerodynamics, collisions, and organic human reactions to the environment.
Gemini 3.8 Flash acts as a guide agent in the real-time loop, observing the scene, breaking down goals into steps, and verifying progress to enable multi-step reasoning and procedural behaviors.
Experiments with Gemini 4 Argon allow the model to break down user-instructed tasks into steps for Genie 4 to render, demonstrated by an Apptronik Apollo humanoid playing checkers.

Currently available as a research preview to trusted testers, with future access planned for Google AI Ultra subscribers (Project Genie) and early access partners in the Gemini API and Google AI Studio.
Known limitations include visual detail drift and object permanence issues in longer sessions or after complex prompt edits.
All generated video and audio includes SynthID, an imperceptible digital watermark for identifying AI-generated content.