Google DeepMind unveils Genie 4 with real-time audio steering

Interactive simulations with native audio and real-time steering

The Genie 4 model generates interactive simulations that react to user inputs instantaneously, incorporating native audio generation, natural movement, and realistic physics. OODA Loop stated that these physics-consistent digital worlds are rendered by the model at 24 frames per second in 720p high-definition.

Users modify active simulations through natural-language prompts to adjust weather, change camera angles, or add new objects and characters. The model generates native audio directly, including ambient sounds like wind and waves, music, and interactive character dialogue that dynamically adapts to simulation events.

This marks the first time a DeepMind real-time world model has generated such sounds, music, and dialogues natively. These features enable creators to build immersive environments through simple text commands.

Architecture powered by Gemini family models

Genie 4 relies on the Gemini model family to function. It uses Gemini Omni to anchor the appearance and behaviors of its simulations.

Ars Technica reported that the model improves upon Genie 3 in "almost every way" by generating experiences frame by frame.

Gemini 3.8 Flash acts as a guiding agent that breaks down goals and evaluates output frames. This ensures the simulation remains on track as users interact with the environment.

Research limitations and safety guardrails

The model is currently an early-stage research project that may encounter detail or object drift during long runs. DeepMind stated that these simulated worlds generally maintain consistency for only a couple of minutes.

TechCrunch reported that users are able to modify the scene by entering prompts during the simulation. Google warned that object permanence and specific details may shift as the simulation progresses.

Google has implemented internal safety guardrails for the project. It uses SynthID watermarking to invisibly label all AI-generated audio and video content.

Research preview and future access plans

Google DeepMind is currently rolling out the model to a limited group of trusted testers. The company is also conducting experiments involving Gemini 4 Argon and Genie 4 in parallel with its Gemini Robotics research.

Future access plans include Google AI Ultra subscribers via Project Genie. Early-access partners will gain entry through the Gemini API and Google AI Studio.