Genie 4
Genie 4 is a real-time interactive world model developed by Google DeepMind and announced on 27 October 2026. It generates explorable environments from text or image prompts that a user can navigate with keyboard and mouse controls while steering the scene with natural-language prompts, and it is the first model in the Genie line to generate audio natively, including ambient sound, music and spoken dialogue from characters in the scene.[1][2][3] DeepMind calls it "our most realistic and interactive world model yet."[2][4] At launch it was available only as a research preview for trusted testers, with access announced as coming to Google AI Ultra subscribers through Project Genie and to early access partners of the Gemini API and Google AI Studio.[1][3]
Genie 4 is the fourth generation of the Genie family, following the original Genie (February 2024), Genie 2 (December 2024) and Genie 3 (August 2025).[7][8][9] Another departure from Genie 3 is a tighter coupling with the Gemini family of models: according to DeepMind, Gemini Omni "anchors the simulations in real-world appearance and behavior," while other Gemini models handle the reasoning needed to keep a simulation coherent over time.[1][2]
Overview
| Attribute | Details |
|---|---|
| Developer | Google DeepMind[1] |
| Announced | 27 October 2026[1][3] |
| Model type | Real-time interactive world model with native audio[1] |
| Predecessor | Genie 3 (announced 5 August 2025)[7] |
| Inputs | Text prompts and reference images; keyboard and mouse for navigation and camera control; text prompts during a session to steer the scene[1][2] |
| Outputs | Video and audio (ambient sound, music, dialogue)[1] |
| Related models | Gemini Omni, Gemini 3.8 Flash, Gemini 4 Argon[1] |
| Launch status | Research preview for trusted testers[1] |
| Announced next access | Google AI Ultra subscribers in Project Genie; early access partners of the Gemini API and Google AI Studio[1][3] |
| Watermarking | SynthID on all generated video and audio[1] |
| Announcement authors | Raia Hadsell, Diego Rivas and Hang Qi[1] |
DeepMind's announcement and model page do not state a resolution, frame rate, parameter count or architecture for Genie 4.[1][2] By comparison, DeepMind specified that Genie 3 ran at 720p and 24 frames per second.[7]
Background: the Genie line
DeepMind's Genie models belong to a research tradition of learned world models, systems that predict how an environment changes in response to actions so that an agent can act inside a simulation. An influential 2018 paper by David Ha and Jürgen Schmidhuber showed an agent trained "entirely inside of its own hallucinated dream generated by its world model."[10] The original Genie paper, posted to arXiv in February 2024, described an 11-billion-parameter "generative interactive environment" trained without action labels on unlabelled Internet videos.[9]
| Model | Announced | What DeepMind said it could do |
|---|---|---|
| Genie | February 2024 (arXiv) | Action-controllable 2D worlds from text, images, photographs or sketches; 11B parameters[9] |
| Genie 2 | 4 December 2024 | Playable 3D environments from a single prompt image; consistent worlds "for up to a minute"[8] |
| Genie 3 | 5 August 2025 | Real-time navigable worlds from a text prompt at 24 fps and 720p, consistent "for a few minutes"[7] |
| Genie 4 | 27 October 2026 | Photorealistic real-time worlds with native audio and dialogue, steerable mid-session with natural language, with Gemini models planning and monitoring the scene[1] |
Genie 3 reached consumers through Project Genie, an "early research prototype" hosted on Google Labs and available to Google AI Ultra subscribers aged 18 and over.[6][11] Thurrott reported that when Project Genie ran on Genie 3, subscribers could explore each world for 60 seconds.[6] The Genie 3 article covers that history.
Capabilities
DeepMind's announcement groups Genie 4's advances into four areas. Its demonstration videos are described as "recordings of real-time interactions" in which users navigate with standard keyboard controls and steer the scene with text prompts; the model page adds that sequences were shortened and are "for illustrative purposes."[1][2]
Real-time steering with natural language
Users can change a running simulation with natural language, for example "changing the weather, moving the camera, or introducing a new person or object." DeepMind says Gemini "will contextually adapt your prompt to match the surrounding scene," and that between user inputs Gemini also "monitors and steers the scene to keep the world progressing cohesively."[1] Prompts can describe both an environment and a character, and some demonstrations start from a reference image used as the first frame.[1][2] TechCrunch described these as "free-form edits" made mid-simulation.[5]
Physics and simulated people
DeepMind says Genie 4 simulates physical scenarios "such as aerodynamics, collisions, and more," and its model page describes environments "where objects collide, water flows, and trees sway in the wind."[1][2] The company describes its simulation of people as "the most dramatic leap": characters that "move naturally, respond organically to their environments, and can be interactively steered in real-time."[1] Demonstration prompts include a crowded 1980s Japanese rural train station, a backstage corridor of a musical theatre and a 1995 Hong Kong film scene.[1][2]
Native audio and dialogue
DeepMind describes Genie 4 as "our first real-time world model to generate audio natively." Each world carries ambient sound for its setting, and the model generates music and interactive dialogue alongside the visuals.[1] DeepMind calls human speech "the true frontier": characters "converse and react to your inputs or events in the world, all in real-time."[1] The model page adds that characters can speak with the user and with each other.[2] The company states that all sound in its demonstration videos was produced directly by Genie 4.[1]
Complex events with Gemini in the loop
For multi-step events, DeepMind runs Gemini 3.8 Flash as "a guiding agent in the loop." While Genie 4 renders the world, Gemini observes the scene, breaks goals into steps, reviews incoming frames to check that each step succeeded, and revises its plan when the world shifts or the user intervenes.[1] TechCrunch summarized the arrangement as Gemini "acting as the director that plans steps and checks frames while Genie creates the visuals."[5]
DeepMind also reported experiments pairing Genie 4 with Gemini 4 Argon as part of its Gemini Robotics research. In one demonstration, Argon breaks a user-instructed task into steps that are passed to Genie 4 as text captions while the simulation renders, producing a simulation of Apptronik's Apollo humanoid playing checkers.[1] The model page describes this as "an infinite playground for robotics simulations."[2]
Availability
At announcement, Genie 4 had been released as a research preview for trusted testers. Google's blog post says it is "coming soon to Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and AI Studio," and DeepMind opened a waitlist for Genie 4 in Project Genie.[1][3] TechCrunch reported the broader rollout as expected "in the coming weeks."[5] DeepMind describes the staged release as intentional: the model was deployed "first to trusted tester partners to get early feedback, before we release the model more broadly."[1] Neither Google's post nor DeepMind's announcement gave a price or a date for the wider rollout.[1][3]
On session length, DeepMind's model page says the model "can support a few minutes of continuous interaction, rather than extended sessions," close to the wording DeepMind used for Genie 3 ("a few minutes of continuous interaction, rather than extended hours").[2][7] TechCrunch reported that Genie 4 "can maintain a consistent scene for up to two minutes, double its predecessor's limit," and Thurrott wrote that Genie 4 will raise Project Genie's per-world limit from 60 seconds to two minutes.[5][6]
Trusted-tester use cases
DeepMind says it worked with cohorts of trusted testers "across creative industries, journalism, education, gaming, and engineering," and reported four kinds of use:[1]
| Area | Use reported by DeepMind |
|---|---|
| Creative direction | A "draft mode" for pre-visualization: filmmakers run multiple "takes" of a scene in real time, adjusting camera angles, lighting, dialogue and action[1] |
| Historical scenarios and photographs | Photojournalists turning a single archival photograph into a simulation of that environment, and visualizing short-term events such as a flash flood or wildfire[1] |
| Interactive experience design | Game developers going from an idea to an explorable prototype, testing visual styles and mechanics before writing engine code[1] |
| Simulation | Enterprise partners evaluating Genie 4 for "zero-to-one" scenario generation[1] |
At a press briefing, Google DeepMind's Diego Rivas said the real-time aspect is most valuable "in the ideation phase, in the storyboarding phase," and said it was not yet clear what the end-to-end product experience would be.[5] The model page also points to training environments for agents such as SIMA.[2]
Hybrid world models
DeepMind presents Genie 4 as an early step toward what it calls "hybrid world models," which "combine the logical, structured reasoning of code and language models with the rich, sensory representation of pixel-level simulation." In this division of labor, Gemini maps structure and plans the narrative timeline while Genie renders "how the world looks, sounds, and responds, pixel by pixel."[1] Tejas Kulkarni, a research director at Google, said during a press briefing: "What we'll increasingly see in the future are these hybrid world models, where you have a combination of Gemini writing code and text, and Genie ingesting that in real time and simulating reality at different levels of abstraction."[5] DeepMind frames world simulation as a step toward artificial general intelligence, writing that training and testing agents in more realistic environments "brings us closer to their successful deployment in the physical world."[2]
Safety and watermarking
DeepMind says Genie 4 "includes strict safety guardrails that are consistent with our safety policies across other models," and that it ran internal safety evaluations before the preview "to prevent the generation of harmful or malicious content."[1] All video and audio generated by Genie 4 is to carry SynthID, Google's imperceptible watermark for identifying AI-generated content.[1][2]
Limitations
DeepMind describes Genie 4 as "an early, evolving technology" and lists the following limitations on its model page:[2]
| Limitation | DeepMind's description |
|---|---|
| Coherent environments | Visual details and object permanence "can shift or drift over time," more noticeably in longer sessions, after multiple complex prompt edits, or when navigating through a scene[2] |
| Other agents | Modeling complex interactions between users, multiple independent agents and shared environments is "an ongoing research challenge"[2] |
| Real-world locations | Genie 4 "is currently unable to simulate real-world locations with perfect accuracy"[2] |
| Text rendering | Clear text "is often only generated when it's in the input world description"[2] |
| Interaction duration | "A few minutes of continuous interaction, rather than extended sessions"[2] |
| Audio accuracy | Real-time audio accuracy "remains an area of ongoing research"[2] |
The blog post adds complex physical interactions and spatial accuracy to the list of open research areas.[1]
Reception and context
TechCrunch placed the launch in a crowded and consolidating world-model field. It noted that Runway had released its GWM Worlds 2 model a couple of months earlier, that World Labs was being acquired by AMD, that Amazon was in talks to buy Decart, and that Rivian had acquired Third Dimension for self-driving simulation, describing Genie 4 as "Google's latest move to stay ahead in the crowded field."[5] Thurrott highlighted the longer session length and real-time prompt edits for Project Genie users.[6]
Team
DeepMind's announcement credits core contributions from a team that includes Hang Qi, Diego Rivas, Aleksander Hołyński, Inbar Mosseri, Jack Parker-Holder, Shlomi Fruchter and Tejas Kulkarni, among many others, and thanks Raia Hadsell, Koray Kavukcuoglu and Demis Hassabis, among others, for guidance.[1]
See also
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39Hadsell, Raia; Rivas, Diego; Qi, Hang. "Genie 4: A new generation of world simulation." Google DeepMind blog, 27 October 2026. deepmind.google/...-generation-of-world-simulation
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22"Genie 4." Google DeepMind model page. Accessed 28 October 2026. deepmind.google/...genie
- ^1 ^2 ^3 ^4 ^5 ^6"We're announcing Genie 4, our new world model for real-time interactive simulation." Google (The Keyword), 27 October 2026. blog.google/...introducing-genie-4
- ^Google DeepMind (@GoogleDeepMind). Post on X, 27 October 2026. x.com/...2115114028338962672
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Bellan, Rebecca. "Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering." TechCrunch, 27 October 2026. techcrunch.com/...ic-worlds-and-real-time-steering
- ^1 ^2 ^3 ^4Giret, Laurent. "Google Makes Genie 4 World Simulation Model Available for Select Partners." Thurrott.com, 27 October 2026. thurrott.com/...odel-available-for-select-partners
- ^1 ^2 ^3 ^4 ^5"Genie 3: A new frontier for world models." Google DeepMind blog, 5 August 2025. deepmind.google/...a-new-frontier-for-world-models
- ^1 ^2"Genie 2: A large-scale foundation world model." Google DeepMind blog, 4 December 2024. deepmind.google/...ge-scale-foundation-world-model
- ^1 ^2 ^3Bruce, Jake; Dennis, Michael; Edwards, Ashley; et al. "Genie: Generative Interactive Environments." arXiv:2402.15391, February 2024. arxiv.org/...2402.15391
- ^Ha, David; Schmidhuber, Jürgen. "World Models." arXiv:1803.10122, March 2018. arxiv.org/...1803.10122
- ^"Project Genie." Google Labs. Accessed 28 October 2026. labs.google/projectgenie
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
1 revision · v1 · 2,143 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Capabilities, availability and quotes checked against Google DeepMind's Genie 4 announcement and model page, Google's Keyword post, and TechCrunch and Thurrott coverage, October 2026.
Cite this page: AI Wiki. "Genie 4." aiwiki.ai, updated 28 Oct 2026, fact-checked 28 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/genie_4