Google DeepMind announces Genie 4 world model for real-time interactive simulation
Google DeepMind announced on October 27, 2026, that it is launching Genie 4, a new-generation world model for real-time interactive simulation. The model is built on the Gemini family, with Gemini Omni anchoring the real-world appearance and behavior of the simulated scenes, while other Gemini models provide complex reasoning to maintain coherence and continuity.
Users can use natural language to move the camera in real time, change the weather, or add new characters and objects while a scene is running. Genie 4 is DeepMind’s first real-time world model that natively generates audio, with each world synchronously generating ambient sound, music, and interactive dialogue. DeepMind has opened a research preview to a small number of trusted testers, and will later roll it out to Google AI Ultra subscribers in Project Genie, as well as to early partners of the Gemini API and AI Studio.
DeepMind said in the announcement that Genie 4 is still in the early research stage. Simulations can remain coherent for short periods, but visual details and object permanence may drift over time, with complex prompt edits or scene navigation; complex physical interaction, real-time audio accuracy, legible text, spatial accuracy, and long-term consistency remain ongoing research areas.
Original sources:
Title: We’re announcing Genie 4, our new world model for real-time interactive simulation. Publisher: Google Publication date: 2026-10-27 URL: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-genie-4/ Nature of source: official announcement Verification status: full text actually read in this report (2026-10-27) Supported content: announcement of Genie 4, integration of Gemini Omni and reasoning models, real-time natural language control, native audio (ambient sound, music, interactive dialogue), opening to trusted testers and subsequent rollout to Project Genie and the API/AI Studio
Title: Genie 4: A new generation of world simulation Publisher: Raia Hadsell, Diego Rivas and Hang Qi / Google DeepMind Publication date: 2026-10-27 URL: https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ Nature of source: official announcement Verification status: key passages and limitation notes actually read in this report (2026-10-27); the page includes demo videos and prompt examples Supported content: natively multimodal synchronized audio generation, Gemini Omni anchoring appearance and behavior, other Gemini models maintaining coherence, real-time natural language scene modification, scope of the research preview, short-period coherence and drift in details/object permanence, and other limitation notes
ABAB AI Insight
The Genie series has been advanced by Google DeepMind since 2024. Genie 3 was announced on August 5, 2025, as the first general world model to support real-time interaction and to generate navigable environments from text or images; Genie 2 before it mainly produced shorter clips. On January 29, 2026, DeepMind connected Genie 3 to the experimental prototype Project Genie, opening it to Google AI Ultra subscribers in the United States and allowing users to create and explore short interactive worlds in a browser. Genie 4 directly uses Gemini Omni for visual and behavioral anchoring, and uses other Gemini models in the loop for multi-step planning and frame-level checks, making it the first in the series to place native audio and a tighter reasoning loop in the same real-time generation pipeline.
In terms of capital and product path, Genie 4 does not involve new financing or equity transactions; it is DeepMind’s model iteration within the existing Gemini and Project Genie channels. Access is clearly layered: first collecting feedback from trusted testers, then Project Genie for AI Ultra subscribers, and early partners for Gemini API and AI Studio. In the announcement, DeepMind positioned world simulation as a step toward AGI, and as a creative tool that can be used for robot and agent safety training environments, as well as film previsualization, game prototyping, education, and manufacturing. The company did not disclose specific session-length limits, resolution, frame rate, or pricing; media reports said the coherent duration is longer than Genie 3, but the official text only emphasizes a “short period” and proactively lists drift limitations.
Compared with Genie 3 in August 2025, the core similarity is real-time interaction and natural-language-driven scene modification; the key difference is that Genie 4 is the first real-time world model to natively generate audio synchronously, including dialogue, and to embed Gemini Omni and iterative reasoning more deeply into the generation process. Other labs were also advancing world models at the same time (for example, some video-generation companies launched models that can respond to input), but DeepMind this time placed audio and the Gemini closed loop in the same product narrative and directly connected it to the existing Project Genie and API channels.
This change mainly improves DeepMind’s ability to handle visual content, physical responses, and sound at the same time in interactive simulation, so future Project Genie and API early users may get scenes with native soundscapes and dialogue. For robot training, the official materials mention experimental directions that combine Gemini models for task decomposition and simulation validation, but the product is still in a research preview, and complex physics and long-term consistency remain open issues. Indicators to watch include when Project Genie is actually opened to AI Ultra users, the real extent of detail drift in sessions, and the usability of native audio and interactive dialogue in feedback from real testers.
ABAB News · Laws of cognition
A world model’s upgrade from “navigable visuals” to a “real-time scene with native soundscapes and dialogue” depends on whether the reasoning model enters the generation loop, rather than simply improving resolution.
When the official text proactively lists drift and audio-accuracy limitations, the accessible scope of a research preview does not automatically mean the product has been stably delivered.
Embedding Gemini-family capabilities into world simulation changes multimodal consistency within the same generation pipeline, rather than adding a separate audio model on its own.