Google DeepMind announced Genie 4 on October 27, 2026, as a real-time interactive world model released as a research preview to trusted testers. It is not a chatbot, and it is not a video generator: it is a world you step into and steer, exploring it as it unfolds around you.
A world you can steer, not just watch
Type a prompt, and a world opens up for you to explore with keyboard and camera controls. Instead of watching a scene play out, you can change the weather, move the camera, or add a person or an object in the middle of the action. Gemini adapts your prompt to the scene, so the world keeps progressing between your inputs rather than waiting for a new command.
Sound, for the first time
Genie 4 is the first real-time DeepMind world model with native audio, so a scene can carry the wind, the waves, and the ambient sound of its setting. It also supports interactive dialogue, with characters that react and talk back in real time, rather than delivering lines on a fixed script.
People, physics, and the brain behind it
DeepMind calls it a new realism standard, with aerodynamics, collisions, and characters that move naturally and respond to their environment, all steerable in real time. Behind it, Gemini 3.8 Flash acts as a guiding agent in the loop: it observes frames, breaks goals into steps, verifies each step, and adapts the plan. Genie 4 has also been tested with Gemini 4 Argon and Gemini Robotics, for example simulating Apptronik's Apollo humanoid playing checkers.
Who is already testing it
Filmmakers are using it for rapid drafts and pre-visualization, trying multiple takes and adjusting camera, lighting, and dialogue live. Photojournalists can turn a single archival photo into a grounded simulation.
Game developers can go from an idea to an explorable prototype in seconds, before writing any engine code. Educators and enterprise testers are generating historical spaces and scenarios like flash floods and wildfires.
The honest limits
It is still an early, experimental model. Details and object permanence can drift over longer sessions, after multiple complex edits, or across wide navigation. Complex physics, real-time audio accuracy, legible text, spatial accuracy, and long-term consistency are still active research areas. All Genie 4 video and audio carries a SynthID watermark. It is in research preview to trusted testers now, and it is coming soon to Google AI Ultra subscribers in Project Genie, plus early-access partners in the Gemini API and Google AI Studio; the Project Genie waitlist is open. It succeeds Genie 3, from August 2025, which ran at 24 frames per second in 720p and held consistency for a few minutes.
Why it matters
For games, film pre-visualization, robot training, and education, the cost of trying out a world drops to a single prompt, powered by models like
Gemini. The question shifts from “can AI draw it?” to “can AI keep a world coherent while you play in it?” Genie 4 is DeepMind's strongest answer yet, even with the drift caveats.
Comments
Post a Comment