Even as Google has moved closer to the AI frontier with Geminii 4 Argon, it’s also iterating on its world model series.
Google DeepMind has announced Genie 4, the newest version of its “world model” that lets people generate and step inside interactive, navigable environments from little more than a text prompt or reference image. The announcement comes just over a year after DeepMind introduced Genie 3, and the new version brings a jump in photorealism, native audio, and far more lifelike human movement.
A world model is a type of AI system built to simulate an environment well enough that a person, or another AI agent, can explore it, interact with it, and have it respond in a physically consistent way. DeepMind sees the category as a core building block on the way to AGI, since an AI that can accurately predict how the world behaves can be trained and tested inside that simulation before ever touching the real one.
What’s new in Genie 4 #
Genie 4 is natively multimodal, meaning it generates matching audio alongside the visuals as a scene plays out — ambient sound, music, and interactive dialogue are all produced in sync with the world rather than layered on afterward. The model draws on Google’s Gemini family to do this: Gemini Omni grounds the simulation’s appearance and behavior in how things look and act in the real world, while other Gemini models handle the reasoning needed to keep a simulation coherent as it unfolds. Users set an event in motion, and Genie 4 works out how the world responds, how it sounds, and how any people in it move, react, and speak.
Google is highlighting four capabilities in particular:
Real-time steering. People can use natural language to reshape a simulation as it runs — changing the weather, adjusting the camera, or dropping in a new object or character — and Gemini adapts the instruction to fit the scene already in progress. In between prompts, Gemini also keeps monitoring the world to hold it together coherently.
Physics and human motion. Genie 4 is built on a deeper understanding of real-world physics, handling things like aerodynamics and collisions more convincingly than before. Google calls the way it simulates people the most dramatic leap of the release — characters move naturally, react to their surroundings, and can be steered interactively in real time, rather than looking like static, scripted figures.
Native audio. Every generated world carries its own ambient soundscape, and characters can hold dialogue that adapts on the fly to what’s happening and what the user types in.
Gemini-driven reasoning for complex events. Pairing Genie 4 with Gemini models adds a planning layer on top of the simulation. Gemini 3.8 Flash acts as a guiding agent, watching the scene as it’s generated, breaking a goal down into steps, and checking incoming frames to confirm each step actually happened — adjusting its plan on the fly if the world shifts or the user intervenes. DeepMind has also been pairing Gemini 4 Argon with Genie 4 and its Gemini Robotics work, using Argon to break a task into steps that are fed to Genie 4 as the simulation renders — demonstrated with Apptronik’s Apollo humanoid robot working through a game of checkers.
Who gets access, and when #
Genie 4 is rolling out as a research preview to trusted testers first, with DeepMind gathering feedback before opening it up more broadly to consumers and developers. The first wave beyond testers will be Google AI Ultra subscribers through Project Genie, alongside early access partners on the Gemini API and Google AI Studio. Interested users can join the waitlist through Google’s sign-up form.
During early testing, DeepMind says teams across filmmaking, journalism, education, gaming, and engineering have been using Genie 4 for things like rapid pre-visualization of scenes, turning a single archival photograph into an explorable historical simulation, prototyping game environments before writing any engine code, and generating quick scenario simulations for enterprise use cases.
What it still can’t do #
DeepMind is upfront that Genie 4 remains an early, experimental model. Simulations can hold together over short sessions, but visual details and object permanence can drift over longer runs, after several complex edits, or while navigating around a scene. Complex physical interactions, accurate real-time audio, legible text rendering, spatial accuracy, and long-term consistency are all listed as areas still under active research.
On the safety side, Google says Genie 4 went through internal evaluations before this preview to guard against harmful or malicious outputs, and that every piece of video and audio it generates carries SynthID, Google’s watermarking system for flagging AI-generated content.
Genie 4 follows a now-familiar pattern for Google’s world-modeling efforts: Genie 1 proved the concept, Genie 3 added real-time interactivity and multi-minute memory, and Genie 4 pushes further into photorealism, sound, and tighter integration with Gemini’s reasoning. DeepMind describes the long-term goal as “hybrid world models” — systems where a language model plans the structure and narrative of a scene while a model like Genie handles the pixel-by-pixel rendering of how it looks, sounds, and responds.