Genie 4: A new generation of world simulation
Google ● Covered by 4 sources
Google DeepMind showed Genie 4, a world model that makes real-time simulations look and sound more real. It can generate audio, dialogue, and motion together, but it’s still only a research preview.
Based on reporting by Google — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind has introduced Genie 4, a new world model built to make real-time simulation look less like a demo and more like a place. The pitch is straightforward: start an event, then let the model carry it forward as the scene reacts, people move, sounds change, and dialogue unfolds at the same time.
The model is natively multimodal, so it can generate music, ambient noise, and interactive speech alongside the world itself. DeepMind says Genie 4 draws on the Gemini family for both appearance and reasoning. Gemini Omni helps anchor the simulation in the look and behavior of the real world, while other Gemini models help keep the whole thing coherent from moment to moment.
DeepMind is rolling it out first as a research preview for trusted testers. Consumer and developer access is still ahead, beginning with Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and Google AI Studio. The company is also using SynthID to watermark video and audio made by the model.
The bigger technical claim is not just prettier scenes. Genie 4 is meant to handle physics, human action, and multi-step events in real time. DeepMind says it can simulate things like aerodynamics and collisions, while also making characters that move naturally, respond to their surroundings, and can be steered on the fly. A guiding agent built with Gemini 3.8 Flash watches the scene, breaks goals into steps, checks progress, and adjusts when the world or the user changes course.
That still doesn’t make it finished. DeepMind is explicit that Genie 4 remains early and experimental, with drift in visual detail and object permanence over longer sessions, especially after multiple prompt edits or more complex navigation. It also says physics complexity, audio accuracy, readable text, spatial accuracy, and long-term consistency are still active research problems. So this is a serious step forward, but it’s not a toy you hand to the public and call it done.
My take — AI-written commentary, not fact-checked reporting
This is classic frontier-model behavior: dazzling on the demo reel, careful in the fine print. The interesting part is not the eye candy; it’s the marriage of pixel soup and planner brains, which is where these systems stop being a parlor trick. Still, a model that needs a safety paragraph this long is not exactly ready to replace anything except bad product hype.
Read more about this at: Google