Genie 4: A new generation of world simulation
DeepMind has announced Genie 4, a world model that generates interactive, real-time simulations with physics, human characters, and native audio including speech. It's in limited research preview with trusted testers, with broader access planned for Google AI Ultra subscribers and API partners.
- Genie 4 generates audio natively alongside the visuals, including ambient sound and interactive dialogue from characters who respond to user input in real time.
- Gemini models act as a guiding layer: Gemini 3.8 Flash watches the scene, breaks goals into steps, checks progress, and adjusts the plan when the user intervenes.
- The article admits limits: visual details and object permanence drift over longer sessions, and text rendering, spatial accuracy, and audio accuracy remain open problems.
Genie 4 is DeepMind's new world model: a system that generates interactive, real-time simulated environments from prompts. The headline changes over the previous versions are native audio (ambient sound plus spoken dialogue generated alongside the visuals), better physics and more natural human characters, and real-time steering through plain-language text. Users navigate with standard keyboard controls and can change weather, move the camera, or add people and objects mid-simulation, with Gemini rewriting those prompts to fit the scene and keeping the world moving coherently between inputs. DeepMind says the model is aimed at building training environments for robots and agents, though the article frames that as a research direction rather than a finished product.
The more interesting technical piece is the Gemini integration. Gemini Omni is said to anchor visual appearance and behavior to the real world, while other Gemini models handle reasoning. In the robotics experiment, a model called Gemini 4 Argon breaks a task into steps and feeds them to Genie 4 as text captions while the simulation renders, so the scene can be adjusted as it goes. Gemini 3.8 Flash runs as a guiding agent in the loop, watching incoming frames, checking whether each step actually succeeded, and replanning when the world shifts or a user intervenes. The article pitches this as a route to "reliable causality" and multi-step reasoning, and sketches a longer-term idea of hybrid world models where language-based planning sits on top of pixel-level simulation. Treat that as a stated direction; the post doesn't show benchmarks for it.
Access is limited for now. Genie 4 is a research preview with trusted testers in creative industries, journalism, education, gaming, and engineering. Google AI Ultra subscribers get it first through Project Genie, and early-access partners of the Gemini API and Google AI Studio come next, with a waitlist open. Early tester uses include filmmakers running multiple takes in a live scene, photojournalists turning archival photos into explorable environments, and game developers prototyping mechanics before writing engine code. Every output carries SynthID watermarking. The post admits visual details and object permanence can drift over longer sessions, and the excerpt cuts off mid-sentence on that limitation, so the full scope of the caveats is unclear from what's available here.