Google DeepMind Opens Genie 4 World Model Research Preview

Google DeepMind has released Genie 4 as a limited research preview, adding native audio, more flexible live scene editing and closer Gemini integration to the latest version of its AI world model.

World models are AI models designed to predict how an environment changes over time, including how objects and surroundings respond to actions. With Genie 4, users can explore these simulated environments using keyboard controls and modify them as they run. They can ask the model to alter the weather, shift the camera or introduce characters and objects without restarting the simulation. Google says Genie 4 also produces ambient sounds, music and responsive dialogue, along with more realistic movement and physical interactions.

Genie 4 can run interactive simulations for up to about two minutes, according to reporting from Google’s launch briefing. However, Google cautions that visual details and object consistency can deteriorate over time, limiting the model’s usefulness for applications that require sustained, accurate simulation.

Gemini Takes a Larger Role in World Simulations

Google is also using Gemini models to help coordinate more complex events within the simulations. The company says Gemini Omni, its multimodal generation and editing model, helps ground scenes in real-world appearance and behavior, while Gemini 3.8 Flash, a model optimized for fast reasoning, monitors incoming frames, breaks goals into steps and adjusts plans as events unfold.

In a separate robotics experiment, Google paired Genie 4 with Gemini 4 Argon, its advanced multimodal reasoning model, to simulate Apptronik’s Apollo humanoid robot playing checkers. Argon broke the task into steps that Genie 4 rendered as the simulation progressed.

Google is exploring this approach to test AI agents and robots in simulated environments before deploying them in the physical world. For now, Genie 4 is available only to trusted testers. Google plans to expand access to Google AI Ultra subscribers through Project Genie and to selected Gemini API and Google AI Studio partners, but did not announce a date.

World Models Draw Investment and Skepticism

Development of world models is attracting substantial investment from across the AI industry. Techstrong’s August special report, “From Words to Worlds,” explores competing world model efforts by Nvidia, Meta, Runway and others. Fei-Fei Li’s World Labs (which AMD agreed to acquire in September) and Yann LeCun’s AMI Labs alone raised more than $2 billion combined earlier this year.

Possible use cases include film studios exploring virtual sets before building them or autonomous vehicle developers simulating dangerous traffic scenarios without putting cars on the road. But the technologies behind these applications vary widely. Some focus on generating realistic visual environments, while others attempt to predict physical behavior. As Techstrong’s Alan Shimel pointed out, the industry has yet to agree on what actually qualifies as a world model.

Beyond the disagreement over definitions, another question stands out: Can these models accurately predict what would happen in the physical world? Techstrong’s report cited the What-If World benchmark, which tested nine video generation models on their ability to simulate how changes in physical conditions affect outcomes. None scored higher than 52%. Google’s Genie models were not evaluated, but the results show how much difficulty the models tested still have with physical cause and effect.

“The gap between ‘looks right’ and ‘behaves like physics’ is still the story,” Shimel wrote.