Google DeepMind releases Genie 4, a world model with audio and real-time text steering

What Genie 4 does

Google DeepMind released Genie 4 on 27 October 2026, the latest version of its world model that generates interactive 3D environments. Unlike earlier generative video models that produce fixed outputs, Genie 4 creates responsive worlds.

The key advancement is native audio. Every world includes ambient sounds, music and interactive dialogue generated on the fly alongside the visual output. Previous Genie versions were silent.

Users steer simulations by typing natural language prompts while they run — moving cameras, changing weather, introducing characters, setting events in motion. The model adapts the world to the description.

TechCrunch described the interactive experience as similar to directing a short film or video game scene through text, with the model handling both the visual generation and real-time synchronized audio on the fly.

The native audio generation is a first for the Genie series. Ambient sounds match the environment type, music adapts to the scene mood, and generated characters speak dialogue that relates to what is happening in the simulation. All audio is produced by the same model that generates the visuals, keeping them synchronized automatically.

How it works

Genie 4 is built on Google DeepMind's Gemini language models, which power real-time steering and dialogue. The model generates photorealistic worlds with advanced physics and natural human motion.

Google says Genie 4 uses Gemini to interpret user text inputs and translate them into scene changes, character movements and audio responses. This is distinct from earlier Genie versions, which were purely video-generation models without language understanding.

World consistency persists across frames. TechCrunch reported Genie 4 produces worlds that evolve rather than resetting on each new input. Characters move through scenes, weather accumulates, and the world state persists.

The Google blog describes Genie 4 as producing the most photorealistic and interactive worlds yet from the Genie family, running at interactive frame rates on Google's infrastructure. The release is limited to trusted testers per TechCrunch.

The release is limited to trusted testers only. Google has not announced a timeline for wider public access. TechCrunch noted the research preview is available to a select group of partners and researchers.

The Genie family

Genie 1 launched in early 2024 as a foundational video model learning game mechanics from internet video without explicit training. Genie 2 followed in December 2024 with improved consistency and playable 3D worlds. Genie 3 added longer persistence and better world coherence.

Genie 4 is the largest jump: adding audio, real-time steering, and significantly improved visual quality. It is the first Genie built on Gemini and the first generating interactive audio.

The research preview approach mirrors how other major AI labs have released generative models cautiously — limited access to trusted testers before wider release.

Genie 4 also brings improved physics simulation compared to its predecessors. The Google blog highlights natural human motion and realistic object interactions as areas where the model has meaningfully improved over Genie 3. Characters can walk, run and interact with objects in the generated world, and the physics engine handles collisions and gravity consistently.

Why it matters

World models are a competitor to game engines and virtual environment builders. If Genie 4 works at the quality claimed, it could let people build interactive 3D experiences from natural language alone, without programming or 3D modeling.

Real-time audio generation with synchronized visual content is technically harder than video-only generation. The audio must match the visual content and respond to user inputs at interactive rates.

Limiting Genie 4 to trusted testers signals Google is aware that real-time generative world models carry misuse risks similar to other generative AI.

The broader implication is that world models are moving from silent, non-interactive video generation toward fully interactive environments that combine vision, language and audio in one system. Genie 4 is the clearest example of that convergence yet from a major lab.

Genie 4 is also a test of whether Google can productize a world model. Competitors including Tencent (Hy-Nomas), Kyutai (CADRE) and others have released open-source world models, but none with audio and real-time steering at Genie 4's reported quality. If Google opens Genie 4 more broadly, it could set the standard for interactive AI-generated environments.

What to watch

Watch whether Genie 4 remains a GEMINI-exclusive test tool or reaches broader availability. The research preview means the public cannot try it yet.

Watch for independent demonstrations validating the real-time steering and audio generation claims, which will separate a genuinely capable model from a polished demo.

Watch whether Genie 4 influences Google's game streaming strategy. An interactive world model could eventually power cloud-based game experiences without the need for traditional game engines, opening a new distribution model for Google.

Watch for the safety evaluation data Google publishes alongside Genie 4. If the model can generate harmful content in real time, the mitigations required will be different from those for static text or image generators.