DeepMind Launches Genie 4 Model with Enhanced Realism in Physics, Humans, and Audio

Image: Ars Technica · Source

DeepMind's new Genie 4 world model generates more realistic physics, human movements, and native audio, enabling interactive, adaptive virtual simulations. It will be available soon via Google's Gemini API and premium AI subscription plans.

Artificial intelligence research lab DeepMind has unveiled Genie 4, the latest evolution of its AI-powered world model technology designed to generate immersive, real-time virtual environments. This successor to the widely noted Genie 3 model introduces significant advances in the realism of simulated physics, human appearances and behaviors, as well as integrated audio, including interactive speech.

Genie 4 leverages DeepMind’s updated Gemini 3.8 Flash reasoning engine to guide the generation of virtual worlds frame by frame based on user prompts. This allows for dynamic content creation where users can alter parameters like weather and camera angles mid-simulation, with Gemini steering these updates to maintain coherence. While previous versions struggled with consistency beyond a few minutes, Genie 4 delivers improved context retention during exploration.

One of the standout enhancements is in the physical realism of simulations. According to DeepMind, Genie 4 models aerodynamic effects and collisions with a degree of natural behavior that surpasses earlier AI-generated video outputs. Similarly, virtual humans now exhibit more lifelike movement patterns and environmental reactions, reducing the uncanny valley effect.

Perhaps most notably, Genie 4 is the first DeepMind world model to incorporate native audio generation. Simulations include ambient sounds and can integrate user-specified audio elements. The model also supports interactive dialog that adapts dynamically to the evolving virtual environment, a leap forward highlighted as a potential game changer for AI video realism.

DeepMind positions these world models as key components in progress toward broader artificial general intelligence (AGI), emphasizing applications such as creating infinite training scenarios for robotics and accelerating prototyping in creative industries. Despite these innovations, the technology remains under active research. Past models like Genie 3 were limited in public availability, and Genie 4 is initially accessible only to select testers.

Google plans to expand access to developers through the Gemini API and AI Studio and offer consumer access via its premium AI Ultra subscription starting at $100 per month. Wider rollout timelines remain unspecified, with phased availability and a waitlist managing early access. Though pricing and usage limits may constrain full potential, Genie 4 heralds a new stage in AI-generated virtual environments with richer sensory and interactive fidelity.