World Models Explained: What They Are & Why 2026 Is Their Year

A world model is a neural network that learns a compressed, interactive simulation of an environment — one an agent can act inside and use to predict what happens next. That single trait separates it from the two things it is most often confused with: an LLM, which predicts the next token of text, and a text-to-video model, which renders a passive, fixed clip. The idea was named by David Ha and Jürgen Schmidhuber in 2018; in 2026 it is suddenly everywhere, because real-time, controllable and persistent world models shipped in barely fourteen months — Google DeepMind's Genie line, NVIDIA Cosmos, and World Labs' Marble. This is an informational landscape guide: DataLLM Lab does not serve world models, so here we explain what they are and where the field actually stands.

What is a world model?

A world model is a neural network that learns a compressed, interactive simulation of an environment — so an agent can predict the future and act inside it. The term was named and formalised by David Ha (then Google Brain) and Jürgen Schmidhuber (NNAISENSE / the Swiss AI Lab IDSIA) in their 2018 paper, presented at NeurIPS as "Recurrent World Models Facilitate Policy Evolution" with an interactive companion at worldmodels.github.io.

Their definition is the one to anchor on: a world model is trained in an unsupervised way to learn a compressed spatial and temporal representation of an agent's environment, so the agent can predict what happens next. The signature result of that paper is the tell for the whole field — Ha and Schmidhuber trained an agent entirely inside its own model-generated "dream" and then transferred that learned policy back to the real environment. If a system can be a place you train an agent, and not just a thing you look at, it is a world model.

Contrast that with a large language model. An LLM predicts the next token of text; it has no built-in notion of an agent moving through space under physics. A world model's entire job is that missing piece — the interactive, physically-consistent substrate an agent acts within.

World model vs LLM vs text-to-video model

The clean line is interactivity plus state: a world model responds to your actions and stays consistent; an LLM predicts text; a video model renders a fixed clip. This table synthesises the distinction from the primary sources — the original 2018 formulation and DeepMind's own framing of Genie versus its Veo video generator.

DeepMind states the video contrast directly: a world model generates the environment frame-by-frame in response to user actions and keeps it physically consistent, whereas a text-to-video generator produces a passive, non-interactive clip. If you are shopping the video side of that table, our best text-to-video guide and the Sora 2 cost breakdown cover it — but keep in mind those are a different category from the world models below.

How a world model works: the perception→model→action loop

It is a closed loop: perceive the environment, compress it, predict the next state, choose an action, feed that action back in. The original architecture makes the parts concrete — three components, each doing one job:

• Vision (a VAE) compresses each raw observation into a small latent vector z — the "what I see, distilled" step.
• Memory (an MDN-RNN) predicts the next latent vector as a probability distribution, given the current latent and the action — the "what happens next" step.
• Controller (a small linear policy) maps the latent plus the memory's hidden state to an action — the "what I do about it" step.

Because the memory model can roll forward on its own predictions, the loop can run without the real environment at all — that is the "dream" the agent trains inside. Modern systems swap the specific networks (Genie 3 generates pixels auto-regressively rather than tiny latents), but the loop is the same shape.

Google DeepMind's Genie line

Genie is the clearest public demonstration of a foundation world model — Genie 2 made worlds playable, Genie 3 made them real-time.

Genie 2, announced by DeepMind on December 4, 2024, is a large-scale foundation world model that generates action-controllable, playable 3D environments from a single prompt image, controlled by a human or an AI agent via keyboard and mouse. Per DeepMind, it can generate consistent worlds for up to a minute, with most shown examples lasting 10–20 seconds. It was released as a research prototype, not a public product. (DeepMind, Genie 2.)

Genie 3 is described by DeepMind as a general-purpose world model and its first real-time, interactive one, generating explorable environments from a text prompt. The verified specs, straight from DeepMind's Genie model page:

• 720p resolution at 20–24 frames per second.
• Environments stay largely consistent for several minutes, with memory recalling changes from specific interactions for up to about a minute.
• Frames are generated auto-regressively, one at a time, conditioned on the initial prompt plus the ongoing sequence of user actions and previously generated frames — which is exactly why it stays interactive and physically consistent instead of producing a fixed clip.
• It supports promptable world events — text-driven changes to weather, objects and characters — and models physical properties like water and lighting.

DeepMind is unusually candid about the limits, and honesty about maturity matters here. Its stated Genie 3 limitations: a limited agent action space, imperfect simulation of other agents, no perfect geographic accuracy for real locations, difficulty rendering readable text, and support for only a few minutes of continuous interaction rather than hours. Genie 3 was first announced in August 2025 as a limited research preview with early access for a small cohort of academics and creators, and is accessible as an experimental "Project Genie" prototype (labs.google) — not a finished commercial product.

The other world simulators: Cosmos and Marble

Two more labs anchor the 2026 landscape, and each takes the idea in a different direction — physical AI, and persistent 3D.

NVIDIA Cosmos is a platform of world foundation models (WFMs) purpose-built for physical AI. It generates physics-based video from combinations of text, image, video and robot sensor/motion data, aimed at use cases like autonomous driving and humanoid robotics. NVIDIA has since launched Cosmos 3, described as an open world foundation model — a "fully open omnimodel" built on a mixture-of-transformers architecture that natively generates text, image, video, ambient sound and actions, with open models downloadable from Hugging Face. (NVIDIA Cosmos; Cosmos 3 announcement.) Treat Cosmos 3's exact launch timing as still settling; the platform and its open-model availability are the durable facts.

World Labs' Marble takes the opposite tack from Genie's streamed frames. Cofounded by Fei-Fei Li, World Labs launched Marble — a multimodal world model — generally available on November 12, 2025. It creates persistent, explorable 3D worlds from text, images, video or coarse 3D layouts. The key difference: Marble's worlds can be explored with no time limit and exported as Gaussian splats, meshes, or videos. That makes the 3D geometry a stored, editable artifact — distinguishing it both from on-the-fly frame-generating world models like Genie and from passive text-to-video generators. (World Labs, Marble.)

Why world models matter: robotics, agents, and games

Because a good world model is a simulator you can train and plan inside — and simulation is the scarce resource in robotics and agents. Three payoffs make the field worth watching:

• Robotics & physical AI. Real-world robot trials are slow, costly and risky. A world model lets a policy practise in generated, physics-aware environments — the explicit goal of NVIDIA Cosmos for autonomous driving and humanoids. It is the "train in a dream, transfer to reality" idea from 2018, scaled up.
• Agents. An agent that can predict the consequences of an action before taking it can plan, not just react. The same instinct that makes a strong LLM agent useful — look ahead, evaluate options — is what a world model supplies for embodied and spatial tasks.
• Games & interactive media. Genie 2 and 3 point at generated, steerable game worlds spun up from a prompt or an image, and Marble at persistent 3D scenes you can export into a pipeline.

The honest caveat runs through all three: today's systems hold consistency for minutes, not hours, and every lab is explicit about fidelity gaps. These are research previews and first releases, not shipped infrastructure.

Why 2026 is their year

Because in barely fourteen months the field crossed three thresholds it had never crossed before — playable, real-time, and persistent. Line up the primary-sourced dates and the acceleration is the story:

Each figure above is drawn from the originating lab's own materials and reflects vendor/lab-reported capability — no independent benchmark scores are implied. What makes 2026 the inflection is not any single release but that interactivity, persistence and physical grounding — the three things a mere video clip lacks — arrived from three different labs at once. That is when a research curiosity becomes a category.

What DataLLM Lab actually serves

To be clear about our own lane: DataLLM Lab does not serve world models or text-to-video models. We are an OpenAI-compatible gateway for chat LLMs and text-to-image models on a single API key. Genie, Cosmos and Marble are not callable through us, and we would not pretend otherwise — this piece is a landscape explainer, not a product pitch.

What is on the gateway that's adjacent to this topic is image generation alongside your chat models. As of July 2026, Google's Gemini image models ("Nano Banana") are callable through the same endpoint — gemini-3-pro-image at $2 / $12 per 1M tokens and gemini-3.1-flash-image at $0.50 / $3 per 1M tokens (input / output; check /pricing for current figures) — as is OpenAI GPT Image. Generative-media pricing moves fast and is often quoted per image or per second of video, so always confirm the unit and date on the source before you budget. Our best AI image API guide covers the image side in depth.

A minimal image call looks exactly like any other request to the gateway — same base URL, same key:

from openai import OpenAI

client = OpenAI(
    base_url="https://www.datallmlab.com/v1",
    api_key="YOUR_DATALLMLAB_KEY",
)

img = client.images.generate(
    model="gemini-3-pro-image",
    prompt="a top-down diagram of an agent acting inside a world model",
)
print(img.data[0].url)

FAQ