Genie 4

Genie 4 is Google DeepMind's most realistic and interactive world model yet, described by the team as a new generation of world simulation. Rather than producing a finished video clip, it builds a world that you can walk through with your keyboard and steer with text while it runs. Text and image prompts generate photorealistic environments, and natural language lets you change those environments live: describe landscapes and events, move the camera, add new characters, or change the weather, then see what happens next. It is positioned for creators, game developers, and AI researchers who want to explore what interactive generated worlds can do, and it is currently offered as a research preview with an open waitlist. Genie 4 addresses a gap that becomes obvious once you compare it with ordinary video generation. A video generator hands you a finished clip: you write a prompt, you wait, and you receive a fixed piece of footage that plays the same way every time. Genie 4 instead builds a world you walk through and steer as it runs, which turns generation into something you can explore and revise rather than something you simply watch. Keeping such a world believable while it is being changed is the hard part. Appearance and behavior have to stay anchored in reality, and the simulation has to hold together from one moment to the next as prompts arrive. Genie 4 is presented as a response to both of those challenges, combining realism with the coherence and continuity a live, interactive world demands. The headline capability is shaping worlds in real time with natural language. You create and change worlds live by describing them: landscapes and events, camera moves, new characters, or a change in the weather. Each instruction is applied while the world is running, so the result is visible immediately as the scene continues to unfold around you. Because the interaction is conversational rather than a one-shot prompt field, you can iterate in the moment: set something in motion, watch how it plays out, then adjust it again. The page pairs this live control with keyboard-driven navigation, so you can move through the generated world yourself instead of only watching a rendered clip. For anyone used to waiting on a batch render, the shift is significant: direction happens continuously, while the simulation is alive and responding. Genie 4 also generates realistic interactive environments rather than static backdrops. In the worlds it produces, objects collide, water flows, and trees sway in the wind. Generated characters do not merely appear; they move and behave naturally within the scene. Those behaviors are what make a world feel inhabitable rather than painted. When you add a character or set an event in motion, the environment responds in ways that read as physical: things bump into each other, liquid moves as liquid, foliage reacts to air. This is the difference between a world you can look at and a world you can test, poke, and move through. It is also what makes the output useful for anyone trying to understand how a scene will behave, not just how it will look in a single frame. Beyond the visuals, Genie 4 generates native audio and live dialogue, so the worlds it builds carry sound as well as imagery, and characters can speak within the simulation as it runs. The model incorporates and builds on capabilities from the Gemini family. Gemini Omni anchors simulations in real-world appearance and behavior, which is how the worlds stay grounded in believable materials and motion. Other Gemini models provide the complex reasoning needed to help maintain a simulation's coherence and continuity over time. That division of labor matters for an interactive world: realism and consistency are separate problems, and both have to be solved simultaneously while a user is actively changing the scene. Bringing native audio and live dialogue into the same system extends that coherence past the visual layer and into how a world sounds and how its characters communicate. Taken together, the workflow Genie 4 describes is simple on the surface and technically demanding underneath. You start from a text prompt or an image prompt, and the model generates a photorealistic world. You then walk through that world with your keyboard while issuing natural-language instructions that reshape it live. Behind the scenes, Gemini models work to keep the simulation grounded and continuous. The page shows this loop in action across a wide range of prompts, and it also links to supporting material through sections on capabilities, future applications, limitations, and responsibility, along with a blog post announcing the model. The emphasis throughout is on interaction as the core of the experience: the world is not delivered to you, it runs with you. The benefits follow directly from that design. Creators get a way to explore an idea spatially, walking through a concept instead of describing it and waiting for a clip. Game developers can see environments in motion, with collisions, flowing water, swaying foliage, and characters behaving naturally, which is closer to how a playable space actually feels. AI researchers get a research preview of a world model they can interrogate interactively, changing conditions on the fly and observing what the simulation does next. Because the model handles realism through Gemini Omni and continuity through other Gemini models, users can focus on direction and observation rather than on manually reconciling a scene that keeps drifting. The result is a faster loop between an idea and a world you can stand inside. The prompts shown on the page illustrate the range of scenarios the model is demonstrated with. A crowned crane walks across snow beneath a clear dusk gradient sky in a natural, documentary style. A fully three-dimensional stained-glass landscape is generated as an explorable crystal world with translucent colored glass, dark lead outlines, and volumetric depth. A first-person, documentary-style skier speeds down an alpine course with motion blur, snow spray, and a subtle game-inspired interface layer. A top-down aerial view frames a still wooden rowboat on a dark pond ringed with lily pads and pink water lilies. Live-action-style footage follows an extravagantly dressed man backstage at an eccentric musical theatre. Other examples include a sunlit meadow with a stream of pink satin fabric winding through it, a watermelon on a stone plinth in slow motion, a first-person hot air balloon view over Antarctica, and underwater wildlife photography of pink axolotls in a clear freshwater river. Across these examples, the same pattern holds: rich environments, natural motion, and the ability to keep exploring. Genie 4 is built for creators, game developers, and AI researchers. Its technical foundation is the Gemini family, including Gemini Omni for grounding simulations in real-world appearance and behavior and other Gemini models for the reasoning that maintains coherence and continuity. Access is currently framed as a research preview, and the page invites visitors to join a waitlist. The site's navigation also flags dedicated sections on future applications, limitations, and responsibility, signaling that the model is being introduced with an explicit account of what it can and cannot yet do. No pricing or plan details are stated in the provided content, and the model is presented primarily as an interactive research release rather than a packaged commercial product. The takeaway is that Genie 4 reframes what a generative world model is for. Instead of handing you a finished clip, it hands you a living world you can walk through, steer with text, and change as it runs, complete with physics, natural character behavior, native audio, and live dialogue, all held together by Gemini. For creators, game developers, and researchers, that means exploration replaces waiting, and a prompt becomes something you can step into and reshape in real time.