Google DeepMind unveils Genie 4 world model: built on Gemini Omni, first with native real-time audio and dialogue, in research preview for trusted testers

Background and pain points

A world model has to generate an environment you can step into and interact with in real time, while keeping physics plausible and the scene consistent. That makes it key infrastructure for training robots and agents, and Google DeepMind calls it an important step toward AGI. On Oct 27 Google DeepMind announced Genie 4 on X and its official blog as its most realistic and interactive world model yet. For now it is only a research preview for trusted testers, not broadly available to consumers or developers.

Architecture and mechanism

Per the blog, Genie 4 builds on the Gemini family: Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models provide the reasoning that keeps a scene coherent. It is natively multimodal and is DeepMind's first real-time world model to generate audio natively, producing ambient sound, music and dialogue alongside the world, with characters that can respond to the user. Users navigate and move the camera with standard keyboard controls and steer the scene with text prompts, such as changing the weather or adding people and objects. For complex events, Gemini 3.8 Flash acts as a guiding agent in the loop: it watches the scene, breaks goals into steps, checks incoming frames to verify each step, and re-plans when the user intervenes. DeepMind also showed Gemini 4 Argon breaking a task into text captions that drive Genie 4 to simulate an Apptronik Apollo humanoid playing checkers.

Benchmarks and verification

Google DeepMind has not published quantitative benchmarks, parameter counts, resolution or frame rate, only demo videos. Stated limitations: objects can drift over longer sessions or after many complex edits, continuous interaction lasts only a few minutes, and text rendering, real-world location accuracy and real-time audio accuracy still need work.

Getting started

Per Google, Genie 4 is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and Google AI Studio; a waitlist is open in Project Genie. All generated video and audio carries a SynthID watermark. For developers the near-term fits are game prototyping, film pre-visualization and simulation environments for robots and agents; API pricing and quotas have not been announced.