Genie 4, a world you can hear

Genie 4, Google DeepMind's real-time world model

Google DeepMind announced Genie 4 on 27 October 2026. It is a world model that builds an explorable scene in real time while a person moves through it and types instructions. Google calls it the lab's first real-time world model to generate audio natively. Sound and speech arrive together with the picture.¹¹1Raia Hadsell, Diego Rivas and Hang Qi, Google DeepMind, "Genie 4: A new generation of world simulation", 27 October 2026: announcement; research preview for trusted testers; release first to Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio; waitlist; tester cohorts from five fields and their uses; text steering and keyboard controls; Gemini 3.8 Flash as guiding agent; Gemini Omni; hybrid world models; native audio and dialogue; the Gemini 4 Argon robotics experiment with Apptronik's Apollo; drift over time; safety guardrails, internal evaluations and SynthID; areas of ongoing research; no resolution, frame rate, latency, parameter count or technical report. https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ Inside the loop, Gemini 3.8 Flash acts as a planner that watches the frames and checks each step.¹¹

Raia Hadsell leads the byline of the announcement. Her co-authors are Diego Rivas and Hang Qi.¹¹ The model page describes Genie 4 as the lab's "most realistic and interactive world model yet".²²2Google DeepMind, Genie model page, "Genie 4", undated, read on 30 October 2026: "most realistic and interactive world model yet"; "coming soon" to Ultra subscribers and early access partners; keyboard and mouse controls; recordings shortened; natively generated audio; the robotics method and the UR5 prompt example; agents such as SIMA; safety text and SynthID; the six limitations, including "a few minutes of continuous interaction, rather than extended sessions". https://deepmind.google/models/genie/ Its tagline, "A new generation of world simulation", also forms part of the title of the post.¹¹²²

Genie 4 is a research preview for trusted testers. Google says it is gathering feedback before a release to consumers and developers.¹¹ That release will start with Google AI Ultra subscribers in Project Genie. Early access partners of the Gemini API and Google AI Studio come in the same first step.¹¹ The model page says Genie 4 "will be coming soon" to those groups and gives no date.²² A waitlist for Genie 4 in Project Genie runs through a Google Form linked from both pages.¹¹²²

Journalism is one of the five fields that supply testers, and engineering is another.¹¹ Filmmakers use a "draft mode" for previsualisation, where they run several "takes" of a scene. Photojournalists turn "a single 2D archival photograph" into a simulation "in seconds".¹¹ Game developers go from an idea to an "explorable prototype in seconds". Enterprise partners use it for what Google calls "zero-to-one" scenario generation.¹¹

From Genie to Genie 3, the earlier world models

Genie 4 is the latest model in a line that began with a research paper. The first Genie appeared on arXiv on 23 February 2024 as a model of "11B parameters".³³3Jake Bruce, Michael Dennis, Ashley Edwards and others, "Genie: Generative Interactive Environments", arXiv 2402.15391, version 1 of 23 February 2024: 11B parameters; training from unlabelled Internet videos; video tokenizer, autoregressive dynamics model and latent action model. https://arxiv.org/abs/2402.15391 It was trained "from unlabelled Internet videos".³³

The paper builds that model from a spatiotemporal video tokenizer, which turns clips into compact tokens, and an autoregressive dynamics model that predicts what comes next.³³ A latent action model infers the action taken between two frames, so actions are learned at the level of single frames.³³

Genie 2 followed on 4 December 2024. Google called it a "foundation world model" that makes "action-controllable, playable 3D environments" from "a single prompt image".⁴⁴4Jack Parker-Holder, Philip Ball, Jake Bruce and others, Google DeepMind, "Genie 2: A large-scale foundation world model", 4 December 2024: playable 3D environments from a single prompt image; keyboard and mouse; autoregressive latent diffusion with a transformer dynamics model; consistency for up to a minute, most examples 10 to 20 seconds; undistilled samples and a distilled real-time version at lower quality. https://deepmind.google/blog/genie-2-a-large-scale-foundation-world-model/ A player steers it with keyboard and mouse. Inside, it is "an autoregressive latent diffusion model" with an autoencoder and "a large transformer dynamics model, trained with a causal mask".⁴⁴

Genie 2 kept a world consistent "for up to a minute". Most of the published examples lasted 10 to 20 seconds.⁴⁴ The demonstration samples came from an "undistilled base model". A distilled version ran in real time, "with a reduction in quality".⁴⁴

Genie 3 arrived on 5 August 2025. It could be navigated in real time at 24 frames per second, at a resolution of 720p. Google said it retained consistency "for a few minutes".⁵⁵5Jack Parker-Holder and Shlomi Fruchter, Google DeepMind, "Genie 3: A new frontier for world models", 5 August 2025: real time at 24 frames per second and 720p; consistency for a few minutes; visual memory of about one minute; emergent consistency; promptable world events; tests with SIMA; limited research preview; its list of limitations, with "extended hours"; no audio generation. https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/ It was the lab's first world model "to allow interaction in real-time".⁵⁵ Its visual memory reached "as far back as one minute ago".⁵⁵

Consistency in Genie 3 was described as "an emergent capability".⁵⁵ The model kept no explicit 3D representation of the scene. Each new frame took into account "the previously generated trajectory", one frame after another.⁵⁵ Genie 3 also introduced what Google called "promptable world events", such as a change of weather.⁵⁵ It went out as a "limited research preview" to "a small cohort of academics and creators".⁵⁵

Project Genie, a web app from Google Labs, opened on 29 January 2026. It ran on Genie 3, with Nano Banana Pro and Gemini working alongside.⁶⁶6Diego Rivas, Elliott Breece and Suz Chambers, Google, "Project Genie: Experimenting with infinite, interactive worlds", 29 January 2026: web app on Genie 3, Nano Banana Pro and Gemini; U.S. Google AI Ultra subscribers aged 18 and over; generations limited to 60 seconds; promptable events not yet included. https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie/ It was limited to Google AI Ultra subscribers in the U.S. aged 18 and over.⁶⁶ Its generations were capped at 60 seconds, and promptable events were "not yet included" in the prototype.⁶⁶

On 19 May 2026 Project Genie gained Street View grounding for places in the U.S. It began a gradual global rollout to eligible Ultra subscribers aged 18 and over.⁷⁷7Diego Rivas, Jonathan Herbert and Nicole Segaran, Google, "Simulate real-world places with Project Genie and Street View", 19 May 2026: Street View grounding for U.S. places; gradual rollout to "all eligible Google AI Ultra $200 subscribers globally (18+)"; Waymo road simulations. https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie-expands/ The same post says Genie helped Waymo "simulate hyper-realistic road environments".⁷⁷

Genie 4 comes with less technical detail than the first Genie. Neither of its two pages gives a parameter count or names the network type.¹¹²² Neither describes the training data or the compute, and no technical report or paper is linked.¹¹²² The gap from the Genie 3 announcement to Genie 4 is 448 days.⁸⁸8Derived from the Genie 3 post (Google DeepMind, 5 August 2025) and the Genie 4 post (Google DeepMind, 27 October 2026): 5 August 2025 to 27 October 2026 is 448 days. https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/ https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/

Steering a running Genie 4 simulation with text prompts

A person can change a Genie 4 world while it runs by typing in plain language. Google's examples include "changing the weather" and "introducing a new person or object".¹¹ Gemini then "will contextually adapt your prompt to match the surrounding scene".¹¹

On the model page the feature appears as creating and changing worlds with natural language, live.²² Gemini keeps working between inputs, and according to the post it "monitors and steers the scene to keep the world progressing cohesively".¹¹

Genie 3 had a similar feature. Its announcement named such edits "promptable world events" and gave similar examples, such as a change of weather or a new character.⁵⁵ Neither Genie 4 page uses that phrase.¹¹²² Project Genie, built on Genie 3, did not offer such events when it opened in January 2026.⁶⁶

The demonstrations pair text with ordinary game controls. The post describes "standard keyboard controls for navigation and camera control".¹¹ The disclaimer on the model page names "standard keyboard and mouse" for the same purpose.²² The difference is small, and both pages agree that text prompts steer the scene.¹¹²²

The model page says all recordings show "real-time interactions from Genie 4". It adds "Results for illustrative purposes" and "Sequences shortened".²² Neither page says how long each recording ran before it was cut.¹¹²²

Gemini 3.8 Flash as the planner inside the Genie 4 loop

Gemini 3.8 Flash works beside Genie 4 as what Google calls "a guiding agent in the loop".¹¹ While Genie 4 draws the world, Gemini "continuously observes the scene" and breaks goals into a sequence of steps.¹¹ It then "reviews incoming frames to track the progress and verify that each step succeeded".¹¹

When the scene shifts or the user steps in, Gemini "adapts its plan in the moment".¹¹ Google credits this loop with "reliable causality" and with "multi-step reasoning" inside a world.¹¹ The model page says the same thing without naming the model: Gemini works "behind the scenes to interpret prompts".²²

Genie 3 worked with an agent too, in a different arrangement. Google tested it with a recent version of its SIMA agent, which sent navigation actions. Genie 3, the post noted, "is not aware of the agent's goal".⁵⁵ In Genie 4 the planner holds the goal and splits it into steps.¹¹

Gemini also has a part in how the world looks. Genie 4, in Google's wording, "incorporates and builds on capabilities from the Gemini family".¹¹ Gemini Omni "anchors the simulations in real-world appearance and behavior". Other Gemini models supply the reasoning that keeps a simulation coherent over time.¹¹ Neither page says whether Gemini Omni is a component inside Genie 4.¹¹²²

Google calls Genie 4 a step toward "hybrid world models", which join the reasoning of code and language models to "pixel-level simulation".¹¹ In the post's words, Gemini "works behind the scenes to map the structure and plan the narrative timeline". Genie meanwhile produces the physical scene "pixel by pixel".¹¹

Rebecca Bellan of TechCrunch reported a briefing with Tejas Kulkarni, whom she described as a research director at Google.⁹⁹9Rebecca Bellan, TechCrunch, "Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering", 27 October 2026: the briefing quote from Tejas Kulkarni; Gemini as "the director"; robot scenarios in simulation; "up to two minutes, double its predecessor's limit", a figure not found on Google's pages; rollout "in the coming weeks"; Diego Rivas on "early days". https://techcrunch.com/2026/10/27/googles-genie-4-world-model-adds-audio-more-realistic-worlds-and-real-time-steering/ In her TechCrunch report, Kulkarni spoke of "a combination of Gemini writing code and text, and Genie ingesting that in real time".⁹⁹ Bellan's own description in the same piece casts Gemini as "the director that plans steps and checks frames while Genie creates the visuals".⁹⁹

How Genie 4 generates sound together with the picture

Genie 4 produces sound at the same time as the images. The post calls Genie 4 the lab's first real-time world model "to generate audio natively". It also calls the model "natively multimodal".¹¹ The page lists music and ambient noise, along with "interactive dialogue", as audio "simultaneously generated with the world".¹¹

Google's text has "audio-like music" at this point, probably a typo for "audio like music".¹¹ Under the demonstrations, Google writes that "All accompanying sounds featured above are produced directly by Genie 4".¹¹ On the model page the line is "Worlds come alive with natively generated native audio", with the doubled word left as published.²²

Speech is the new part. Google writes that Genie 4 "introduces interactive dialogue that adapts on the fly".¹¹ Characters "converse and react to your inputs or events in the world, all in real-time".¹¹ The model page adds that characters can speak with the user and with each other.²²

The Genie 3 announcement of August 2025 mentions no audio generation.⁵⁵ That fits the claim that Genie 4 is the first real-time world model from the lab to generate audio natively.¹¹⁵⁵

Google marks everything Genie 4 generates. All video and audio from the model "will include SynthID", and the watermark covers "visual and audio modalities".¹¹²² Sound is also on the list of open problems. The model page says "Real-time audio accuracy remains an area of ongoing research and improvement".²²

A robotics test with Gemini 4 Argon and Genie 4

Google has paired Genie 4 with Gemini 4 Argon to explore simulations for robots.¹¹ The post describes the work as an experiment that brings frontier research "together with Gemini Robotics".¹¹ In the setup, "Argon breaks down a user-instructed task into steps for Genie to render".¹¹

The steps travel as text. They "are interleaved as text captions to Genie 4 as the simulation is rendering".¹¹ The model page describes the same method: Argon "creates a detailed multi-step plan", and its steps reach Genie 4 as captions.²² The demonstration shows Apptronik's Apollo humanoid "playing a game of checkers", with the task broken "into real-time steps".¹¹

Argon became generally available in the Gemini API on 13 October 2026.¹⁰¹⁰10Google, Gemini API release notes (changelog), entry of 13 October 2026: "Gemini 4 Argon (gemini-4-argon) generally available (GA)", as cited in the PARALLAX model card for Gemini 4 Argon. https://ai.google.dev/gemini-api/docs/changelog Genie 4 was announced 14 days later.¹¹¹¹11Derived from the Gemini API release notes (Google, 13 October 2026) and the Genie 4 post (Google DeepMind, 27 October 2026): 27 minus 13 is 14 days. https://ai.google.dev/gemini-api/docs/changelog https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ PARALLAX covered the launch of that model in an earlier explainer on Gemini 4 Argon.

The model page shows a second robot in a prompt example. A "Universal Robots UR5" arm holds a pipette under the heading "Direct complex events".²² That example illustrates a prompt. Google does not present it as an experiment.²²

Google gives no numbers for the robotics work. Neither page reports a success rate or a quantitative evaluation.¹¹²² Neither says whether any plan was tried on a physical robot.¹¹²² The model page calls the setup "an infinite playground for robotics simulations".²² TechCrunch's Rebecca Bellan wrote that the experiment "shows a possible way to test robot scenarios in simulation before they happen in the real world".⁹⁹

Agents are the other stated use. The model page imagines worlds in which agents such as SIMA "can learn new skills".²² It argues that testing agents in "truer-to-life environments" brings their use "in the physical world" closer.²² The post calls simulating the world "a crucial aspect of building AGI".¹¹

How long a Genie 4 world stays consistent

Google gives no number of seconds for how long a Genie 4 world holds together. The model page says only that it "can support a few minutes of continuous interaction, rather than extended sessions".²² Neither page states a memory horizon in seconds or minutes.¹¹²²

The post describes the drift in words. Genie 4 "can maintain coherent environments over short sessions".¹¹ Even so, "visual details and object permanence can still shift or drift over time".¹¹ On the model page, drift grows "as a simulation runs longer". It also follows "multiple complex prompt edits" and movement through a scene.²²

Genie 2 held a world "for up to a minute".⁴⁴ Genie 3 stayed consistent for "a few minutes", the same wording now used for Genie 4. Its visual memory reached back about one minute.⁵⁵²² Project Genie capped each generation at 60 seconds.⁶⁶

A different figure appeared in the press. Rebecca Bellan wrote in TechCrunch that Genie 4 "can maintain a consistent scene for up to two minutes, double its predecessor's limit".⁹⁹ That figure appears on neither Google page, and it conflicts with Google's own wording of "a few minutes".¹¹²² Two minutes is 120 seconds, twice the 60 second cap of Project Genie.¹²¹²12Derived from the TechCrunch report (Rebecca Bellan, 27 October 2026), the Project Genie post (Google, 29 January 2026) and the Genie 3 post (Google DeepMind, 5 August 2025): 2 minutes is 120 seconds, and 120 divided by the 60 second cap is 2; Genie 3's stated horizon is "a few minutes", with no number. https://techcrunch.com/2026/10/27/googles-genie-4-world-model-adds-audio-more-realistic-worlds-and-real-time-steering/ https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie/ https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/ Genie 3's announcement claimed "a few minutes", so the doubling in the TechCrunch report seems to be measured against the Project Genie cap.⁵⁵¹²¹²

Safety, limitations and access for Genie 4

Google says Genie 4 ships with "strict safety guardrails" in line with the policies for its other models.¹¹²² Before the preview, the lab ran "internal rigorous safety evaluations to prevent the generation of harmful or malicious content".¹¹²² Both pages say the model went first to trusted tester partners on purpose, to gather feedback before a wider release.¹¹²²

The model page opens its list of limitations with "Genie 4 is an early, evolving technology".²² The list has six entries. One is "Coherent environments", about drift. Another is "Interaction and simulation of other agents", which the page calls "an ongoing research challenge".²² A third admits the model is "currently unable to simulate real-world locations with perfect accuracy".²²

Text rendering has its own entry. "Clear and legible text is often only generated when it's in the input world description", according to the page.²² The duration entry repeats the "few minutes" limit, and the last entry covers audio accuracy.²² The post adds "complex physical interactions" and "spatial accuracy" to the areas of ongoing research.¹¹

Much of this list was already there for Genie 3. Its announcement named the same entries for other agents and for real-world locations. It also named text rendering and interaction duration.⁵⁵ Genie 4 adds "Audio accuracy" and "Coherent environments" and drops Genie 3's "Limited action space".²²⁵⁵ The duration entry changed its last words from "rather than extended hours" to "rather than extended sessions".⁵⁵²²

Access is still closed to the public. Google AI Ultra subscribers in Project Genie come first, along with early access partners of the Gemini API and AI Studio.¹¹²² Neither page gives a price for Genie 4 or a list of countries.¹¹²² Neither sets an age limit.¹¹²²

Project Genie itself offers some context. Since 19 May 2026 it has been rolling out to "all eligible Google AI Ultra $200 subscribers globally (18+)".⁷⁷ The post does not say whether the $200 is a monthly figure.⁷⁷ Rebecca Bellan of TechCrunch wrote that Genie 4 would reach those groups "in the coming weeks".⁹⁹ Google's own pages say only "coming soon".²²

Diego Rivas of Google DeepMind spoke to TechCrunch as well. Bellan reported that he "stressed that it's still early days for Genie 4". She added that, in his view, the end-to-end product experience is not yet clear.⁹⁹

What Google has not disclosed about Genie 4

Google has published no technical specifications for Genie 4. Neither page states the resolution or the frame rate, and neither gives a figure for latency.¹¹²² For Genie 3 the lab had published 720p at 24 frames per second.⁵⁵ No session cap or memory horizon is given in seconds.²²

The model itself is undescribed. There is no parameter count and no network type.¹¹²² The pages name no training data and no compute, and link no paper or technical report.¹¹²² The first Genie, by contrast, was published as a paper that gave its size, 11 billion parameters.³³

The coupling between the two models is also left open. Neither page says how often Gemini 3.8 Flash reviews frames.¹¹²² Neither gives a latency budget for the loop, and the role of Gemini Omni is unspecified.¹¹²²

The robotics demonstration carries no metrics and no result on a physical robot.¹¹²² The safety evaluation is described in one sentence, with no categories or results.¹¹²² Neither page mentions red teaming by outside groups.¹¹²²

Access has no date beyond "coming soon".²² There is no price and no model identifier for the Gemini API.¹¹²²

Genie 2 came 285 days after the first Genie paper, and Genie 3 followed 244 days later.¹³¹³13Derived from the dates of the Genie paper (arXiv, 23 February 2024), the Genie 2 post (4 December 2024), the Genie 3 post (5 August 2025), the Project Genie posts (29 January 2026 and 19 May 2026) and the Genie 4 post (27 October 2026): the gaps are 285, 244, 177, 110 and 161 days, and their sum is 977 days. https://arxiv.org/abs/2402.15391 https://deepmind.google/blog/genie-2-a-large-scale-foundation-world-model/ https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/ https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie/ https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie-expands/ https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ Project Genie opened 177 days after Genie 3. Its expansion came 110 days after that, and Genie 4 arrived 161 days later.¹³¹³ From the first Genie paper to Genie 4 is 977 days.¹³¹³ For Genie 4, the dates of the next stages are not disclosed.²²