Google DeepMind releases Genie 4 with real-time audio and reasoning
Genie 4 integrates multiple Gemini models to simulate complex physics and audio, allowing users to modify environments in real-time. The release highlights DeepMind's strategy of combining disparate architectures to create interactive AI systems.
Google DeepMind has released Genie 4, a world model that integrates Gemini Omni for visual anchoring, Gemini 3.8 Flash for physical reasoning, and dedicated audio models for generative sound. Unlike previous iterations, Genie 4 allows users to modify the simulation in real-time via live prompts, such as changing weather conditions or introducing new objects. The model simulates complex physics, including aerodynamics and collisions, and features interactive characters that move naturally. It supports two-minute simulations, double the limit of Genie 3, and is currently available to Gemini Ultra subscribers.
The release demonstrates a shift toward combining disparate model architectures to produce novel AI systems. Google demonstrated a robotics application where Gemini 4 Argon breaks down tasks into steps for Genie 4 to render, enabling robot training in simulated environments. Logan Kilpatrick reported that Gemini 4 Argon processed 25 trillion tokens in its first 72 hours. The company is also preparing Gemini 4 Carbon, an ultra-large base model designed for rapid post-training iteration. TPU sales have grown significantly, with new customers including OpenAI and Oracle.