---
title: "DeepMind’s Genie 4 Doubles World-Model Memory to Two Minutes, Adds Audio"
description: "The world model that crashed gaming stocks in January now talks, and still forgets its own worlds after 120 seconds"
author: "Luis Goa"
published: 2026-10-27T16:00:22Z
modified: 2026-10-27T18:41:09Z
url: https://rews.cc/a/deepmind-s-genie-4-doubles-world-model-memory-to-two-minutes-914a3e
language: en
tags: ["ai", "deepmind", "world-models", "robotics", "gemini", "tech"]
publisher: "Rews (https://rews.cc)"
---

# DeepMind’s Genie 4 Doubles World-Model Memory to Two Minutes, Adds Audio

*The world model that crashed gaming stocks in January now talks, and still forgets its own worlds after 120 seconds*

By Luis Goa · October 27, 2026 · https://rews.cc/a/deepmind-s-genie-4-doubles-world-model-memory-to-two-minutes-914a3e

## In brief

- DeepMind’s Genie 4 sustains a consistent simulated world for about 2 minutes, double Genie 3’s roughly 1 minute
- Genie 4 adds native audio (ambient sound, music, dialogue) generated alongside each video frame
- The model uses Gemini 3.8 Flash as a planning layer that evaluates frames, plus “Gemini Omni” for real-world grounding
- Genie 3’s January 2026 public launch wiped out roughly $15bn in gaming-stock market value in one trading day, including a 24% drop in Unity
- Genie 4 is a limited research preview; wider access via Project Genie and the Gemini API is planned for “the coming weeks”

Genie 4 forgets what it built after about two minutes. That is the headline number in Google DeepMind’s research preview, released October 27, and it is also the ceiling on what the model can currently be used for: a world model is a system that generates a navigable 3D environment frame by frame, from a text prompt, with no pre-built level geometry and no programmed physics engine underneath it. You type what you want, the model renders the next frame based on what you typed and what you did a moment before, and it keeps doing that for as long as it can keep the world consistent with itself.

Two minutes is the latest point on a short, steep curve. DeepMind’s original Genie, announced in 2024, generated 2D platform-game worlds at a single frame per second with no meaningful persistence. Genie 2, released that December, held a 3D world together for [10 to 20 seconds in most of DeepMind’s own demonstrations](https://deepmind.google/blog/genie-2-a-large-scale-foundation-world-model/), with an outer limit the company put at up to a minute. Genie 3, unveiled in [August 2025 on DeepMind’s blog](https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/), rendered at 720p and 24 frames per second and sustained a coherent world for roughly a minute. Genie 4 doubles that to two minutes, according to [TechCrunch’s reporting on the launch](https://techcrunch.com/2026/10/27/googles-genie-4-world-model-adds-audio-more-realistic-worlds-and-real-time-steering/), the first independent figure to pin the gain to a specific number rather than DeepMind’s own framing of “improves on Genie 3 in almost every way,” as [Ars Technica put it](https://arstechnica.com/google/2026/10/deepminds-genie-4-world-model-adds-audio-and-more-realistic-people-to-simulations/).

The doubling matters because consistency, not resolution or frame rate, is the hard part of this problem. DeepMind’s own description of Genie 3 noted that generating an environment autoregressively, frame by frame, is harder than generating an ordinary video because small errors compound: a wall painted in one direction has to still be painted in that direction two minutes later, from an angle the model has never explicitly stored. There is no 3D mesh sitting in memory the way there would be in a game engine built on Unity or Unreal; the world exists only as a sequence of generated frames and whatever the model has managed to carry forward from them. DeepMind told TechCrunch that details and objects can still drift the longer a simulation runs, and that the team has not solved how to keep in-world audio accurate or in-world text legible over time.

## What’s actually new

The feature DeepMind is leading with is not the longer memory window but native audio: Genie 4 generates ambient sound, music and interactive dialogue in the same pass as the visual frames, rather than bolting audio on afterward. Users can still redirect a running simulation in plain language, mid-scene, changing weather, camera position or adding characters, a capability Genie 3 already had. What changes the mechanics underneath is the role Gemini now plays. Genie 4 calls on Gemini as what DeepMind describes to TechCrunch as a planning layer: one Gemini model breaks a user’s goal into steps and plans what should happen next, while Genie generates the pixels, and a separate component, which DeepMind calls Gemini Omni, grounds the simulated world’s appearance and behavior against real-world data. Ars Technica’s reporting names the planning component specifically as Gemini 3.8 Flash, which evaluates each output frame to keep the simulation from drifting off its original goal.

**How Genie 4 generates a world** Division of labor between the planning model and the world-generation model

| From | To | How |
| --- | --- | --- |
| User prompt / mid-scene edit | Gemini 3.8 Flash *(plans steps, evaluates each frame)* | goal |
| Gemini 3.8 Flash *(plans steps, evaluates each frame)* | Genie 4 core model *(generates frame, audio, physics)* | next step |
| Gemini Omni *(grounds appearance in real-world data)* | Genie 4 core model *(generates frame, audio, physics)* | grounding |
| Genie 4 core model *(generates frame, audio, physics)* | Rendered world *(drifts after ~2 minutes)* | frame + audio |
| Rendered world *(drifts after ~2 minutes)* | Gemini 3.8 Flash *(plans steps, evaluates each frame)* | frame review |

Based on [TechCrunch and Ars Technica reporting on DeepMind's October 27, 2026 announcement](https://techcrunch.com/2026/10/27/googles-genie-4-world-model-adds-audio-more-realistic-worlds-and-real-time-steering/)

That architecture is close to how a reasoning model breaks a hard question into intermediate steps before committing to an answer, except here the intermediate steps are frames of a simulated world rather than tokens of text. The same mechanism also feeds Google’s robotics work: TechCrunch reports that a Gemini 4 Argon model, the frontier model Google [put on general sale at $19.99 a month](https://rews.cc/a/google-puts-gemini-4-argon-on-sale-the-frontier-model-now-co-400db7) earlier in October, was used to break robot tasks into steps that Genie 4 then simulated in real time, as a way to generate training scenarios for physical robots without needing a real robot or a real room.

## The stock-market precedent

Genie’s commercial release carries a track record. When Google rolled Genie 3 out to AI Ultra subscribers through Project Genie on January 29 and 30, 2026, gaming stocks fell within a single trading day: Unity Software dropped 24%, erasing roughly $5 billion of market value, according to [Bloomberg’s reporting at the time](https://www.bloomberg.com/news/articles/2026-01-30/unity-video-game-stocks-fall-as-google-s-ai-tool-genie-sparks-fears). Roblox fell 13%, losing about $3 billion; Take-Two Interactive fell 7.9%; CD Projekt fell 8.7%; Ubisoft fell roughly 7%. The gaming sector lost an estimated $15 billion in combined market value in that one session on fears that a model which can generate a playable environment from a text prompt might make traditional game engines unnecessary. Unity, Roblox and Take-Two had each recovered at least 3% of the loss by the following Monday, once analysts pointed out that Genie 3’s one-minute consistency window, lack of multiplayer and lack of precise, scripted physics left it nowhere near replacing a production engine for an actual shipped game.

Genie 4 arrives with that episode as the market’s reference point rather than a surprise, which is probably why the October 27 announcement has not reproduced January’s selloff. The model is not yet broadly available: DeepMind has opened it to a limited group of trusted testers, with access planned, in the company’s words, “in the coming weeks,” for Google AI Ultra subscribers through Project Genie and for early-access partners using the Gemini API and Google AI Studio, per [Google’s own announcement](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-genie-4/). No price has been disclosed for Genie 4 access specifically, separate from the existing AI Ultra subscription.

Genie 4 is not the only world model competing for this niche. Runway shipped GWM Worlds 2, a playable world model with its own audio generation, months before Genie 4, according to TechCrunch’s reporting, and the surrounding infrastructure is consolidating: AMD has acquired Fei-Fei Li’s World Labs, Amazon is reportedly negotiating to buy world-model startup Decart, and Rivian has acquired Third Dimension for self-driving simulation work. DeepMind’s research director Tejas Kulkarni told TechCrunch the company expects “hybrid world models,” combining a reasoning layer like Gemini with a generative layer like Genie, to become the standard architecture, which is a reasonable bet given that Genie 4 already needs Gemini to keep its own worlds from falling apart.

The number to watch next is not two minutes, it’s whatever DeepMind reports for Genie 5. The jump from Genie 2 to Genie 3 was threefold; Genie 2 to Genie 4 doubled it once more. If DeepMind holds even a 1.5x pace per release, the next version clears five minutes of consistency, which is roughly the length of a single level in a simple game, and the point at which “world model” starts to mean something closer to what alarmed Unity’s shareholders in January.

## See also

- [Genie (AI model) — Wikipedia](https://en.wikipedia.org/wiki/Genie_%28AI_model%29) — en.wikipedia.org · Version history and technical specs across Genie 1 through 3