DeepMind announces Genie 4
On 27 October 2026 Google DeepMind announced Genie 4, a world model that generates an interactive world in real time together with sound: ambient noise, music and characters' speech. The world can be changed with text prompts as it runs, and Gemini models watch the scene and steer it step by step. It is a research preview for trusted testers; for Google AI Ultra subscribers in Project Genie and early-access partners in the Gemini API and Google AI Studio it is "coming soon", with a waitlist.
Why it matters
Genie 3 (August 2025) gave a world a viewer could move through in real time. DeepMind says Genie 4 is its first real-time world model to generate audio natively, and pairs it with Gemini as a guide that breaks a goal into steps and checks the frames. It is a shift from a world that is shown to a scene that is steered in conversation. A reader cannot try it yet: the announcement describes a preview, not a release.
What the post says. The model is "natively multimodal": ambient noise, music and dialogue are generated simultaneously with the world. Gemini Omni anchors the simulation in real-world appearance and behaviour, and other Gemini models supply the reasoning that keeps it coherent. Four capabilities are named: changing the world in natural language as it runs (weather, camera, a new person or object); realism of physics (aerodynamics, collisions) and of human movement; its own audio, including dialogue that adapts to events; and Gemini 3.8 Flash as a "guiding agent in the loop" that observes the scene, breaks goals into steps and checks that each step succeeded. In the robotics example Gemini 4 Argon breaks a task into steps that are passed to Genie 4 as text captions; the demonstration is a simulation of Apptronik's Apollo humanoid playing checkers. Testers in film, photojournalism, games and enterprise describe drafting a scene, turning an archival photograph into an environment and prototyping a game. Access. A research preview for trusted testers; "coming soon" for Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio; there is a waitlist. All video and audio output carries SynthID. Limits, from the model page. Visual detail and object permanence can drift, especially in longer sessions and after several complex edits; accurate real-world places, legible text (generally only when it is in the world description), interaction among several agents and audio accuracy are open problems; the model supports "a few minutes" of continuous interaction, not long sessions. What the record does not claim. The post gives no frame rate, resolution or session length, and names Genie 3 only in the site's navigation: it makes no comparison with it. "First real-time world model to generate audio natively" is DeepMind's own wording; no independent confirmation was found on 28 October, and the post appeared the day before this record. The video and sound on the page were not played. What the testers report is the company's own account.