DeepMind Releases Genie 4: Natural Language Becomes the Remote Control for Real-Time Worlds

Google DeepMind today released Genie 4. This is the next generation of the Genie family of world models, focused on real-time interaction and native audio. Whereas past video generation models could only output a closed clip, Genie 4 turns generation into a simulated world you can enter and modify.

In the official demo, users control navigation and the camera with the keyboard, and use text prompts to change the scene: change the weather, move the camera, and add people or objects to the scene. For example, make a pink silk ribbon appear on a grassy field, or place the camera inside the basket of a hot-air balloon above the Antarctic ice sheet. The scene responds in real time, and characters continue to act according to the new setup.

This capability is powered by models in the Gemini family. Gemini Omni is responsible for making the visuals closely match real-world appearance and behavior, while other Gemini models translate requests into actions, plan complex events, observe the frames, and track progress. When the user changes instructions, Gemini adjusts the plan on the fly.

Genie 4 is also the first real-time world model in the Genie series to natively generate audio. Ambient sound, music, and character dialogue are all produced in sync with the scene. In the demo, someone flies on a dragon’s back, and the wind, wingbeats, and fire sounds all come from the model itself, not from sound effects added in post-production.

DeepMind’s official blog breaks this update down into four areas: real-time scene control, physics and character motion, sound modeling, and complex event simulation. The physics work involves aerodynamics and collisions, and characters can walk, observe, and react to the environment naturally. In the complex-events part, Gemini 3.8 Flash acts as a guiding agent, observing the frames being generated in real time, breaking goals into steps, and checking whether each step is completed. In another experiment, Gemini 4 Argon and Genie 4 work together to let Apptronik’s Apollo humanoid robot play checkers in simulation.

On safety, Genie 4 comes with strict safety guardrails, and all generated video and audio carry SynthID watermarks. The team also acknowledges current limitations: after running for a long time, visual details and object consistency drift, and complex physics, real-time audio accuracy, text rendering, and spatial accuracy are all still under research.

Genie 4 is currently available as a research preview to trusted testers. Next, it will open to Google AI Ultra subscribers in Project Genie, and early partners through the Gemini API and Google AI Studio will also get access. The official waitlist is already open; those who want to try it early can sign up here: https://forms.gle/WQnSNEM3omF7Uruu7.

Developer Ethan Cross commented on X: real-time interactive world simulation, combined with multimodal control and synchronized sound, is the real frontier. Some users also posted screenshots of a login prompt, asking why it can only be used through Google AI Ultra. The bigger question for many is: when will ordinary users actually be able to play with it? For now, the answer is probably: get in line first.