DeepMind's new world model holds a consistent scene for two minutes, double its predecessor, with improved physics and real-time user steering.
Google's Genie 4 world model now generates interactive environments with audio, more realistic physics, and humanlike characters, while letting users steer scenes in real time. The model maintains coherent scenes for up to two minutes—twice the limit of Genie 3—marking a meaningful jump in generative world simulation.
The release signals Google's push to commercialize world models beyond research demos, positioning the technology for gaming, robotics training, and simulation-based enterprise applications.
World models that sustain longer, controllable scenes move the technology closer to production use in synthetic training data and virtual prototyping, a market enterprise buyers in robotics and gaming should start evaluating now. Google is signaling it intends to own this layer before competitors catch up on scene persistence.
The daily signal, curated. Get it in your inbox.
Subscribe on LinkedIn →