Genie 4: A new generation of world simulation
Google DeepMind announced Genie 4, a new world model it describes as delivering improved realism and real-time interactive simulation. The announcement is credited to Raia Hadsell, Diego Rivas and Hang Qi. Genie 4 is being rolled out first as a research preview for trusted testers, with the company gathering feedback before broadening access. The stated path to wider release starts with Google AI Ultra subscribers inside Project Genie, alongside early access partners using the Gemini API and Google AI Studio. A waitlist for Genie 4 in Project Genie is open.
The core claim is that Genie 4 is natively multimodal: it generates music-like audio, ambient noise and interactive dialogue at the same time as it generates the world itself. DeepMind says all accompanying sounds in its demonstrations were produced directly by the model. This is the first real-time world model from the team to generate audio natively, according to the announcement, with each simulated environment carrying the ambient sound of its setting, from wind in trees to waves on a beach. Human speech is framed as the harder frontier: characters hold conversations and react to user inputs or in-world events as they happen.
Genie 4 builds on the Gemini family. Gemini Omni is used to anchor simulations in real-world appearance and behavior, while other Gemini models supply reasoning intended to keep a simulation coherent and continuous. In practice, the user can set an event in motion and the model simulates what follows: how the environment responds, how it sounds, and how people in it move, react and speak. Users steer scenes with text prompts and navigate with standard keyboard controls for movement and camera.
Four capabilities are highlighted. The first is real-time shaping of a simulation through natural language: changing weather, moving the camera, or introducing a person or object, with Gemini adapting the prompt to the surrounding scene. Between user inputs, Gemini also monitors and steers the scene to keep the world progressing cohesively. The demonstration prompts are unusually detailed, specifying film stock, grain, contrast, lighting, camera behavior and audio, which suggests the model is meant to honor fine-grained art direction rather than just coarse scene descriptions.
The second is physics and human action. DeepMind says Genie 4 is built on an advanced physical understanding of the world and simulates scenarios such as aerodynamics and collisions. The headline claim here is how it handles people: characters that move naturally, respond to their surroundings and can be steered interactively in real time. The third is audio, covered above. The fourth is complex, multi-step events handled with Gemini reasoning. Gemini 3.8 Flash acts as a guiding agent in the loop: as Genie 4 renders, it observes the scene, decomposes goals into steps, reviews incoming frames to check progress, and revises its plan when the world shifts or the user intervenes. DeepMind frames this as enabling reliable causality, procedural behaviors and multi-step reasoning.
DeepMind also describes experiments pairing Gemini 4 Argon with Genie 4 for Gemini Robotics. Argon breaks a user-instructed task into steps that are fed to Genie 4 as interleaved text captions while the simulation renders, allowing the scene to be steered toward task completion. One example simulates Apptronik's Apollo humanoid playing checkers. The stated aim is simulating training environments for robots and agents, and DeepMind situates world simulation within its broader work toward AGI, emphasizing safe, infinitely variable simulations.
Trusted testers span creative industries, journalism, education, gaming and engineering. DeepMind reports that testers highlighted the jump in photorealism, native audio and real-time steering as enabling rapid concept-to-simulation work; creatives are described as using Genie 4 as a fast draft mode. The company is explicit that Genie 4 remains an early, experimental model, and that the tester program is intended to reveal where interactive world models fit real workflows and where they still need to grow.
A few things are notably unspecified: pricing beyond the Ultra subscription tier, latency and frame-rate targets, resolution, maximum session length, and how consistency holds over long interactions. DeepMind characterizes real-time performance as a core feature but gives no benchmark figures. This suggests would-be adopters should treat the preview as a way to probe stability and controllability rather than a production service.
A likely trade-off is the pairing with Gemini agents. Routing scene generation, monitoring and step-by-step planning through separate Gemini models could improve causal consistency and multi-step tasks, but it also implies layered inference cost and added latency between a user's input and the world's response, which matters for anything billed as real-time. Developers should also expect the steering interface to be prompt-driven rather than a conventional game engine API, which changes how content pipelines and QA are organized.
Why it matters: World models that generate photorealistic, steerable environments with synchronized audio and dialogue point toward cheaper synthetic training grounds for robotics and agents, and a new authoring surface for games and interactive media. Access is currently gated behind trusted testers, an Ultra subscription and partner programs, so most developers cannot yet build on Genie 4 directly; the practical near-term step is watching how the Gemini API and AI Studio early access evolves. For teams in simulation-heavy work such as robotics evaluation, the Argon-plus-Genie experiments are the most relevant signal, though the experimental status and missing performance figures mean any reliance on it remains speculative.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Genie 4: our new world model unlocking unprecedented realism in real-time simulation.