DeepMind’s Genie 4 brings audio and sharper realism to AI-generated worlds

Google DeepMind has a new world model, and the gaming industry is probably checking its portfolio again. Genie 4 improves on Genie 3 in almost every way, according to Ars Technica’s reporting. The upgrades in focus are audio and more realistic simulations. What Genie 4 changes A world model generates a virtual environment on the fly instead of loading pre-built scenery. Nothing is rendered ahead of time. The system produces what you see in response to what you do. Genie 3 already worked this way, with clear limits. It launched in August 2025 with real-time interactive output at 720p and 24 frames per second. It also had a one-minute memory limit, so the world it built could start losing track of itself fairly quickly. Genie 4 is aimed at those rough edges. DeepMind researchers had previously discussed better realism, longer memory and audio as features for future versions of the model. Sound and realism are the improvements now attached to the new release. Realism has also been climbing on the Genie 3 side. As of late October 2026, Project Genie integrated Google Street View data to produce photorealistic street-level environments, with a broader rollout planned. How we got here Th...