Google introduces cutting-edge Genie 4 world model
SiliconANGLE Maria Deutscher ● Covered by 6 sources
Google previewed Genie 4, a model that builds 3D worlds from text. It can also add sound and edit scenes without wrecking the realism.
Based on reporting by SiliconANGLE, Maria Deutscher — 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 has previewed Genie 4, the newest version of its world model line, and this one is built to do more than sketch a scene. It can generate realistic three-dimensional environments, add audio like crashing waves, and handle moving characters with noticeably better fluid motion than its predecessor.
The pitch is simple enough: describe a world in natural language, then step into it. Users can steer the simulation, tweak the weather, add objects, and move the camera with either prompts or keyboard controls. Google DeepMind staffers Raia Hadsell, Diego Rivas and Hang Qi say the model’s biggest jump is in how it simulates people, making them move naturally and react to what’s around them in real time.
Genie 4 doesn’t do that alone. Google says it leans on Gemini Omni for realistic visual assets and Gemini 3.8 Flash to keep individual frames consistent. According to the DeepMind team, Gemini watches the scene as it changes, breaks tasks into steps, and adjusts when the world shifts or the user interferes. That matters because consistency is the whole game here. If the model drifts, the illusion falls apart.
DeepMind is also looking at robotics. In that setup, Genie 4 is paired with Gemini 4 Argon so engineers can simulate robot behavior from high-level instructions. A short prompt, like a robotic arm soldering two car parts, can be expanded into a long sequence of detailed prompts for each step of the task. Google says more and more precise prompting improves rendering quality, which makes the simulations more useful for training robots from visual examples. For now, Genie 4 is being tested with a limited group of early adopters, with access planned through Google AI Ultra, the Gemini API and Google AI Studio.
My take — AI-written commentary, not fact-checked reporting
This is Google doing what Google does best: turning a research demo into a productized pipeline before anyone else finishes the press release. The interesting part isn’t the shiny world model; it’s the plumbing underneath, where cheap reasoning, video generation and editing get welded together into something useful for robot training. Open models will keep bragging about freedom, but the closed shops are clearly winning the race to turn compute into something that looks like a workflow.
Read more about this at: SiliconANGLE