Google DeepMind announces Genie 4, a world model with sound and live editing
The model generates ambient sound and speech with the scene and takes text edits mid-simulation; access is limited to trusted testers, and the post gives no resolution, frame rate or consistency duration.
- Models & capabilities
- Notable
Google DeepMind announced Genie 4, a successor to Genie 3 that it calls a “natively multimodal” world model. Where Genie 3 produced explorable 3D worlds, Genie 4 generates music, ambient sound and speech along with the visuals. The announcement, by Raia Hadsell, Diego Rivas and Hang Qi, describes it as DeepMind’s “first real-time world model to generate audio natively”, with dialogue that adapts to what the user does.
Users move through a running simulation with keyboard controls and can change it with text prompts mid-session, for instance altering the weather, moving the camera or adding a person or object. DeepMind says a Gemini model adapts those prompts to the existing scene and keeps the world progressing between inputs. For multi-step tasks, a Gemini model observes the scene, breaks a goal into steps, checks whether each succeeded and revises the plan; a separate demo pairs Gemini 4 Argon with Genie 4 to simulate a humanoid robot playing checkers. The post says Gemini Omni anchors the simulations in real-world appearance and behaviour. DeepMind positions the model for robot and agent training environments as well as creative and game prototyping, and lists early-tester uses including pre-visualising scenes and turning a single archival photograph into an environment.
The post gives no resolution, frame rate or consistency duration, and is explicit about limits: visual detail and object permanence can drift, especially in long sessions or after several prompt edits, and complex physical interactions, text legibility and audio accuracy are listed as active research areas. Some secondary summaries cited a two-minute consistency figure; the post itself does not state one.
Genie 4 is a research preview for trusted testers, with a waitlist. DeepMind says rollout is planned next to Google AI Ultra subscribers in Project Genie and then to early-access partners through the Gemini API and Google AI Studio. All generated video and audio will carry SynthID watermarks, and the company says the model went through internal safety evaluations before the preview.