Google DeepMind announces Genie 4 world model with native audio and real-time language steering

Google DeepMind released Genie 4 on October 27, 2026, a new world model designed for real-time interactive simulation with native audio generation.

Core capabilities

The model generates photorealistic environments, simulates physics such as aerodynamics and collisions, and produces ambient sounds, music, and interactive dialogue directly from the simulation state. Users steer scenes through natural language prompts—changing weather, camera position, or introducing characters—while the system maintains coherence using supporting Gemini models.

Genie 4 builds on Gemini Omni for real-world anchoring and other Gemini components for reasoning and continuity. A guiding agent based on Gemini 3.8 Flash breaks down goals into steps, reviews frames, and adapts plans in real time.

Rollout and access

The model is currently available as a research preview to trusted testers. It will reach Google AI Ultra subscribers through Project Genie and early access partners via the Gemini API and Google AI Studio. A waitlist is open for Project Genie.

Intended uses and limitations

Trusted testers in creative industries, journalism, education, gaming, and engineering are exploring it for pre-visualization, historical scenario reconstruction, game prototyping, and scenario generation. All generated video and audio includes SynthID watermarking.

Current limitations include drift in visual details and object permanence over longer sessions, challenges with complex physical interactions, spatial accuracy, and legible text rendering. DeepMind positions the preview as a step toward hybrid world models that combine pixel-level simulation with structured reasoning from language models.

Safety evaluations preceded the limited release, and guardrails align with existing model policies. The company emphasizes gathering feedback before broader availability.

Sources

• deepmind.google
• blog.google

Cover photo by SHVETS production on Pexels, used under the Pexels License.