Google DeepMind launches Genie 4 world model

Google DeepMind has launched Genie 4, a research preview world model for real-time world simulation with native audio. The model, unveiled on October 27, 2026, is now rolling out to a limited group of trusted testers. It generates photorealistic and interactive virtual environments with synchronized audio and real-time scene editing.

Genie 4 is described as DeepMind's first real-time world model to generate audio natively. Each world contains ambient sound and interactive dialogue in real time. The model enables real-time control through natural language prompts to modify weather, camera angles, and add objects while maintaining scene consistency for up to two minutes at 24fps 720p resolution.

Technical Capabilities

Genie 4 renders explorable, physics-consistent digital worlds. It models real-world physics scenarios including aerodynamics and collisions. Characters move naturally and respond organically to environments.

The model allows freeform edits during simulation. Users can type prompts mid-simulation to change the scene. It maintains a consistent scene for up to two minutes, which is twice as long as its predecessor, Genie 3. Genie 4 also improves physics and creates more human-like characters compared to the previous version.

Potential applications for Genie 4 include games, interactive learning, robotics, and film prototyping. It is natively multimodal, generating music, ambient noises, interactive dialogue, photorealistic visuals, fluid movement, and native audio simultaneously.

Architectural Integration

Genie 4 builds on the Gemini family of models. Gemini Omni anchors simulations in real-world appearance and behavior. Other Gemini models provide complex reasoning for coherence and continuity.

Specifically, Gemini 3.8 Flash acts as a guiding agent integrated with Genie 4. It breaks down goals into steps, reviews each output frame, and adapts plans to maintain reliable causality and multi-step reasoning. In one confirmed experiment, a Gemini 4 Argon model broke down tasks into steps that Genie 4 simulated in real time during a robotics test with an Apptronik Apollo humanoid playing checkers.

DeepMind plans future hybrid world models that combine logical reasoning from code and language models with Genie's pixel-level sensory rendering. Research director Tejas Kulkarni stated that these future models would combine Gemini writing code or text with Genie simulating reality.

Practical Applications and Limitations

Trusted testers currently evaluating Genie 4 include filmmakers using it for rapid pre-visualization and ideation. Photojournalists are creating grounded simulations from 2D photographs. Game developers are prototyping ideas in seconds. Enterprise partners are evaluating the model for rapid zero-to-one scenario generation and for identifying gaps in simulation capabilities.

Google DeepMind engineer Diego Rivas said real-time simulation is valuable for rapid ideation in storyboarding and film prototyping.

Despite its advances, Genie 4 has acknowledged limitations. Details and objects can drift or lose stability during longer simulations. The model faces challenges with object permanence, audio accuracy, and text legibility. Complex physical interactions, spatial accuracy, and long-term consistency beyond a couple of minutes remain areas for improvement. Google warns that as simulations run longer, details and object permanence can shift over time.

The model includes strict safety guardrails and uses SynthID, DeepMind's imperceptible digital watermarking technology, to identify AI-generated content. Following this limited preview, Genie 4 will be available to Google AI Ultra subscribers in Project Genie and early access partners via the Gemini API and AI Studio.