Google DeepMind's latest release pushes that boundary further than any version before it, and the results have caught significant attention from both the research community and creative industries.

World simulation has been steadily advancing from a research curiosity into something that looks and feels uncomfortably close to real life.

Google DeepMind Debuts Genie 4 World Model

Google DeepMind announced Genie 4, its latest generation of interactive world model, describing it as unlocking unprecedented realism in real-time simulation. The model is natively multimodal, meaning it generates audio simultaneously alongside the visual world, including ambient sounds, music, and interactive dialogue, all produced in a single unified pass rather than layered on afterward.

Genie 4 is built on a foundation that incorporates capabilities from two Gemini models. Gemini Omni grounds the simulations in real-world appearance and behavior, while other Gemini models provide the reasoning needed to maintain coherence and continuity as a simulation unfolds in real time.

The architecture allows a user to set an event in motion and have Genie 4 simulate what happens next across visuals, sound, and the actions of people within the scene.

The announcement was authored by Raia Hadsell, Diego Rivas, and Hang Qi of Google DeepMind, and the team framed the model as an advance toward what the company sees as a core challenge in building AGI: the ability to understand and simulate the physical world in a safe, infinitely variable way.

Genie 4 World Model: More Realistic Humans, Physics

The most dramatic single improvement in Genie 4 compared to prior versions is how the model handles people. The research team described human simulation as the area where Genie 4 makes its biggest leap, with characters that now move naturally, respond organically to their environments, and can be interactively steered in real time rather than following scripted or rigid animation paths.

Alongside human motion, Genie 4 introduces what the team describes as a new standard of realism for physical simulations. The model can simulate real-world physics scenarios including aerodynamics, collisions, object interactions, and material behavior, all processed in real time as the user interacts with the world.

Demonstrated examples from the announcement include a watermelon shattering on a stone plinth with airborne droplets visible in slow motion, a maple seed falling naturally through the air, and a crowded 1980s Japanese train station populated with dense, naturally moving crowds.

Real-time steering through natural language is another core capability. Users can modify a simulation mid-session by typing prompts to change the weather, move the camera, or introduce a new person or object. Gemini contextually adapts those prompts to match the surrounding scene, so changes feel integrated rather than abrupt. Between user inputs, Gemini monitors and steers the world to keep it progressing cohesively without needing constant input.

Massive Improvements on New Video Generations

Audio is the other area of significant change in Genie 4. Every generated world now carries ambient sound matching its setting, whether that is wind moving through trees, waves on a beach, or the background noise of a busy indoor environment. Google DeepMind described this as the first real-time world model to generate audio natively, meaning sound is not post-processed or added from a library but is generated simultaneously with the visual content.

The deeper audio frontier the team highlights is interactive human dialogue. Characters in a Genie 4 simulation can now converse with each other and react to user inputs or environmental events in real time, adapting what they say and how they say it as the scene changes around them. This is distinct from a pre-scripted voice line being triggered by a condition. The dialogue emerges from the same real-time simulation as the rest of the world.

For complex multi-step scenarios, Genie 4 works in tandem with Gemini 3.8 Flash as an active guide. As Genie 4 generates a world in real time, Gemini continuously observes the scene, breaks goals into sequential steps, reviews incoming frames to verify each step succeeded, and adapts the plan if the user intervenes or the world changes unexpectedly. This pairing was demonstrated in simulations of a robotics task, where Gemini 4 Argon and Genie 4 together simulated a humanoid robot playing a game of checkers, with Argon breaking the task into real-time steps that Genie rendered frame by frame.

The team acknowledged that Genie 4 still has meaningful limitations. Visual details and object permanence can shift during longer sessions, and complex physical interactions, spatial accuracy, rendering legible text, and long-term consistency remain active areas of research. All video and audio generated by Genie 4 will carry SynthID, Google DeepMind's digital watermarking system, to identify the content as AI-generated.

Frequently Asked Questions

What Is Google DeepMind?

Google DeepMind is an AI research laboratory formed in 2023 from the merger of Google Brain and DeepMind, both of which were previously separate divisions. The organization is part of Alphabet, Google's parent company, and functions as the primary research arm responsible for developing frontier AI systems across a wide range of domains including language models, scientific discovery tools, robotics, and generative media.

Its most recognized contributions include the AlphaFold protein structure prediction system, which won a Nobel Prize in Chemistry in 2024, the AlphaGo and AlphaZero game-playing systems, and the Gemini family of multimodal AI models. Google DeepMind is led by Demis Hassabis, who co-founded the original DeepMind in London in 2010 before it was acquired by Google in 2014.

What Is the Genie World Models?

The Genie series is Google DeepMind's line of interactive world models, designed to generate and simulate environments that users can explore and interact with in real time rather than simply watch as a fixed video. The first public generation, Genie 2, was released in late 2024 and demonstrated the ability to generate 3D-explorable environments from a single image.

Genie 3 followed in August 2025, advancing photorealism and interactive control. Each generation in the series builds on a core idea: rather than producing a single fixed video clip, the model generates a continuous, responsive simulation that reacts to user inputs, camera movements, and natural language prompts as the session unfolds. Genie 4, released as a research preview in October 2026, adds native audio generation, substantially more realistic human motion, and real-time dialogue to that foundation, making it the most capable version yet.

Who Can Access the Genie 4 World Model?

Genie 4 launched as a research preview and is not yet available to the general public. Google DeepMind has been rolling it out first to trusted testers across creative industries including filmmaking, journalism, education, game development, and engineering in order to gather structured feedback before a broader release.

Following that initial phase, access will expand to Google AI Ultra subscribers through a feature called Project Genie, and to early access partners of the Gemini API and Google AI Studio. There is no confirmed pricing for Genie 4 access beyond the existing Google AI Ultra subscription tier, and no specific date for when the general consumer rollout or developer API access will begin. People interested in gaining early access can join the official Genie 4 waitlist through the form linked in Google DeepMind's announcement page.