Skip to content
Neural Daily

ModelsNews

Genie 4: a real-time world model with audio and dialogue

🇭🇺 Olvasd el magyarul →

Google DeepMind has unveiled Genie 4, a world model that generates photorealistic, steerable worlds in real time with native audio and dialogue. For now, only trusted testers have access.

By: Mikola Máté Patrik · · 2 min read

Google DeepMind cover image for Genie 4: a hummingbird feeding on an orange flower inside a glass sphere, beside the title text
Image: Google DeepMind · Press kit · Official cover image of the Google DeepMind blog post · Source

What was announced?

Google DeepMind announced Genie 4, the successor to Genie 3 (Google DeepMind). The model generates real-time, interactive world simulations: users navigate and move the camera with the keyboard and steer the scene with text prompts as it evolves. Genie 4 is natively multimodal and generates music, ambient sound and dialogue together with the world.

What changed since Genie 3?

  • Real-time steering in natural language: weather, camera, or a new person or object mid-simulation. Gemini adapts the prompt to the scene and keeps the world coherent between inputs.
  • More realistic physics (aerodynamics, collisions) and what DeepMind calls the biggest leap: human characters that move naturally and can be steered.
  • DeepMind's first real-time world model with native audio, with characters that talk and react in real time.
  • Gemini 3.8 Flash runs as an agent in the loop: it breaks goals into steps, watches incoming frames and checks whether each step succeeded.

DeepMind is also experimenting with Gemini 4 Argon: in one demo, Apptronik's Apollo humanoid plays checkers in simulation while Argon breaks the task into real-time steps.

Who can try it, and when?

It is currently a research preview for trusted testers. It is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio. A waitlist for Project Genie is open. All generated video and audio carries a SynthID watermark.

Limitations

DeepMind itself calls it an early, experimental model. It stays coherent over short sessions, but details and object permanence can drift in longer runs or after several complex prompt edits. Complex physical interactions, audio accuracy, legible text and long-term consistency remain open problems. The announcement gives no resolution, frame rate or benchmark figures.

Why it matters

World models are moving toward simulation environments for training robots and agents, and wiring in Gemini as a planner makes that more concrete. If you build games, prototypes or simulations, watch the Gemini API early access, since that is where Genie 4 will become something you can actually integrate.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles