⏱️ Reading time: 9 min
Google DeepMind has just unveiled Genie 4, a model capable of simulating complete worlds that respond in real time to the actions of whoever controls them. It is the fourth generation of its world model, the family of systems the company uses to build interactive environments from scratch.
📑 En este artículo
The central novelty isn’t just the imagery. Genie 4 natively generates ambient audio, music, and dialogue while building the scene. DeepMind presents it as a step toward something bigger: a necessary step toward building artificial general intelligence (AGI), since understanding and simulating the physical world is, according to the company, a key ingredient for achieving it.
TL;DR
- Google DeepMind unveiled Genie 4, a world model with audio and video generated in real time.
- The system combines Gemini Omni for appearance with other Gemini models for coherence.
- For now, it’s a research preview limited to selected trusted testers.
- Google AI Ultra subscribers in Project Genie and Gemini API partners will be added later.
- Users control the camera with the keyboard and modify the scene with text prompts.
What Is Genie 4
Genie 4 is Google DeepMind’s world model, a system that generates real-time interactive simulations by natively combining video, ambient audio, and dialogue. Built on the Gemini family, it lets a user trigger an event and have the system respond with coherent imagery, sound, and behavior.
Unlike an end-to-end generated video, Genie 4 functions as an environment, since it responds to the actions of whoever controls it. The user navigates with the standard arrow keys while a text prompt can change the weather, move the camera, or introduce a new object into the scene. DeepMind describes it as the foundation for creating safe, infinitely variable simulations, useful both for training robots and agents and for building new interactive experiences.
What Happened
DeepMind announced Genie 4 on October 27, 2026 as a research preview, not a public launch but a testing phase with trusted users before opening access to a wider circle. The official announcement explains that the company is gathering feedback from that small group before moving on to consumers and developers.
The access plan has two entry points. The first is Project Genie, the product where DeepMind has been testing this model family, which Google AI Ultra subscribers will access first. The second is an early access program for Gemini API and Google AI Studio partners, aimed at developers who want to integrate these simulations into their own tools.
According to the official blog, the system lets you set an initial event, for example a character playing saxophone on a floating platform in the middle of a lake, and lets Genie 4 decide how it continues: how the environment responds, how it sounds, and how the people within it move and talk.
Background and History
Genie 4 doesn’t appear out of nowhere. It’s the fourth installment in the Genie family, which DeepMind has been developing since 2024. The first version generated playable two-dimensional worlds from a single image; Genie 2 made the leap to more consistent three-dimensional environments; Genie 3 added real-time interaction with standard keyboard and camera controls. Each generation solved a different problem, such as spatial persistence, simulation duration, or responsiveness to user actions.
What was missing, and the leap that defines this generation, was sound and dialogue generated alongside the imagery. Through the previous version, DeepMind’s worlds were silent: you could move through them, but you couldn’t hear them. Generating audio coherent with a moving image is, from a design standpoint, a different problem than generating video alone, because it requires synchronizing two modalities that evolve over time while maintaining scene coherence. That’s why DeepMind presents it as a full generation, not as an incremental update to Genie 3.
| Version | What It Added | Limitation |
|---|---|---|
| Genie (2024) | Playable 2D worlds generated from a single image | No 3D persistence or audio |
| Genie 2 | More consistent, longer-lasting 3D environments | No dialogue or ambient sound |
| Genie 3 | Real-time interaction with keyboard and camera | Silent worlds, no native audio |
| Genie 4 | Natively generated audio, music, and dialogue | Access limited to trusted testers |
How Genie 4’s World Model Works
The mechanism behind Genie 4 combines two types of models from the Gemini family, each with a distinct job. Gemini Omni anchors the simulation in realistic appearance and behavior. It decides how light looks on water or how an arm moves while playing an instrument. The rest of the Gemini models handle the complex reasoning needed to maintain the scene’s coherence and continuity over time.
That division of labor explains why Genie 4 doesn’t lose track of what happened thirty seconds earlier in the simulation. While one model generates the appearance frame by frame, the other oversees that the physics, objects, and characters stay consistent from one image to the next, something generative video models without a reasoning component tend to lose over time.
User control has two channels. One is physical: the standard arrow keys to navigate and move the camera, just like in a video game. The other is natural language, through a text prompt that can change the weather, reposition the camera, or introduce a new character, with Gemini adapting that request to blend with what already exists in the scene. Between one instruction and the next, the same system keeps watching the simulation so it progresses coherently without additional intervention.
DeepMind’s own announcement illustrates this with a concrete example: a scene of a person playing saxophone on a floating platform in a lake, with fog, diffuse light, and the instrument’s sound generated alongside the image. It’s the kind of atmospheric detail that used to require an entire production team and that can now be described in a text prompt.
💭 Key takeaway: Genie 4 doesn’t generate a video that it then plays back. It generates frame by frame, in response to what the user does, like a video game engine that invents its own content on the fly.
There’s no direct way to verify from the outside how the work is split between Gemini Omni and the reasoning models, because DeepMind didn’t publish architecture details, weights, or evaluation metrics alongside the announcement. For now, the only available source is the video demonstration accompanying the official blog.
flowchart TD
U["User: keyboard and text prompts"] --> G4["Genie 4"]
G4 --> GO["Gemini Omni: appearance and behavior"]
G4 --> GR["Gemini models: reasoning and coherence"]
GO --> W["Simulated world: video, audio, and dialogue"]
GR --> WImpact and Analysis
DeepMind frames this work within its stated goal of advancing toward AGI. Understanding and simulating the world, the company says in its announcement, is a crucial aspect of achieving that. The bet is that a realistic world simulator can serve as a training environment for robots and agents, a testing ground for behaviors without the cost or risk of doing so in the physical world.
That application isn’t new in the industry. Various AI labs have been exploring simulated worlds as a source of training data for robotics, precisely because gathering real data of a robot interacting with its environment is slow and expensive. What distinguishes this world model is that it targets full multimodal simulations, sound and dialogue included, rather than limiting itself to geometry and visual physics.
The other use DeepMind explicitly mentions is creating new interactive experiences, closer to entertainment than to robotics. There, the open question is one of quality and control. A simulation that invents what comes next can pleasantly surprise or drift from what the user expected, and that balance between creativity and predictability is exactly what the current preview with trusted testers aims to measure before opening access.
Opening up a system that generates realistic people, voices, and environments raises safety questions that go beyond technical performance. What it can be asked to simulate, how abusive uses get filtered out, and what happens to consent when the model generates a recognizable person are all open questions. DeepMind hasn’t publicly detailed its safeguards for Genie 4, beyond limiting initial access to a small group of testers.
⚠️ Heads up: Genie 4 still isn’t available to the general public or to most developers. It’s a closed preview, and DeepMind didn’t publish any performance metrics, maximum session length, or resolution alongside the announcement.
What’s Next
Access will open in layers. First, it continues with the trusted testers already using it so DeepMind can gather feedback. After that, the company plans to add Google AI Ultra subscribers within Project Genie and early access partners of the Gemini API and Google AI Studio. DeepMind hasn’t published a date for general availability or for end consumers.
It also didn’t detail pricing terms, usage limits, or content policies for when the system leaves its closed testing circle. These are reasonable questions for a model that generates complete worlds with realistic people, voices, and sounds. How much control does DeepMind retain over what the model can simulate? And what happens the day any user, not just a selected tester, can write the initial prompt?
Try it yourself: sign up for the Project Genie waitlist on DeepMind’s official blog to see if you get early access to Genie 4.
Frequently Asked Questions
What Is Google DeepMind’s Genie 4?
Genie 4 is the fourth generation of DeepMind’s world model, a system that generates real-time interactive simulations with native video, audio, and dialogue, built on the Gemini family.
How Is Genie 4 Different from Genie 3?
Genie 3 already let you navigate real-time generated worlds with keyboard and camera, but those worlds were silent. Genie 4 adds ambient audio, music, and dialogue generated alongside the image, not added afterward.
Who Can Test Genie 4 Right Now?
Only a group of trusted testers selected by DeepMind. Broader access will arrive later for Google AI Ultra subscribers in Project Genie and for early access partners of the Gemini API.
Does Genie 4 Generate Real Audio or Just Video?
It generates both natively and simultaneously. Music, ambient noise, and interactive dialogue are produced alongside the images, not added later with a separate audio model.
What Is Genie 4’s World Simulator For?
DeepMind positions it as a training environment for robots and agents, and as a foundation for creating new interactive experiences, since simulating the world is, according to the company, a key step toward AGI.
When Will the World Simulator Be Available to Developers?
DeepMind hasn’t given a date for general availability. Early access for developers will arrive via the Gemini API and Google AI Studio, after the current closed testing phase.
References
- Google DeepMind: official announcement of Genie 4 and its capabilities.
- Google AI for Developers: Gemini API documentation for early access.
- Wikipedia: general context on generative AI models.
- Google DeepMind: official site of the research lab.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Featured image: Foto de Steve A Johnson en Unsplash
Did it work for you? Got a different error? Say so below: questions get answered and help the next reader.
Leave a comment
0 Comments