Fei-Fei Li’s World Labs has just pulled back the curtain on Atlas, an omni world model that represents one of the most ambitious attempts yet to teach artificial intelligence how to understand and generate the physical world in three dimensions. Announced in early access on September 1, Atlas is not simply another large language model bolted onto a video generator — it is, according to the company, a system built from the ground up to think spatially.
The implications are sweeping. From robotics to film production, from autonomous driving to architectural design, a model that natively reasons in 3D rather than treating space as a post-processing problem could reshape entire industries.
What Is Atlas?
Atlas is what World Labs calls an “omni world model” — a single system pretrained from scratch to operate across text, images, video, and 3D simultaneously. Unlike traditional AI pipelines that handle each modality separately and stitch the results together, Atlas combines all inputs into a shared spatial context. From that context, it generates what comes next, maintaining 3D consistency with everything it has already processed and extrapolating what lies beyond the visible frame.
Technically, Atlas is a multimodal autoregressive diffusion transformer. The autoregressive component means it generates output sequentially, predicting each token based on all prior context. The diffusion component handles the probabilistic, high-fidelity generation of images and video. Together, they allow Atlas to operate at 1440p resolution and produce video sequences of up to one minute in length — substantially higher quality than most competing systems.
In head-to-head evaluations, human raters preferred Atlas over rival video generation models in 75 to 94 percent of trials, depending on which competitor was tested. While benchmark comparisons in AI are always contested terrain, the margin here is notable.
The Road to Atlas: From Marble to Spatial Intelligence
Atlas is not World Labs’ first public model. The company released Marble in November 2025, a commercial multimodal model that enabled 3D world creation from various input types. Marble demonstrated the core architecture and attracted significant enterprise interest, but Atlas represents a generational leap in capability.
Where Marble was a proof of concept demonstrating that spatial reasoning and generative modeling could coexist in a single system, Atlas is a mature production model targeting real-world deployment. The jump in resolution, temporal coherence, and cross-modal consistency between the two models reflects over a year of intensive training runs and architectural refinements.
Fei-Fei Li’s Long Bet on Spatial AI
World Labs was founded in February 2024 by Fei-Fei Li, who previously directed Stanford University’s AI Lab and served as Chief Scientist of AI/ML at Google Cloud. Li has spent much of her career arguing that vision and spatial intelligence are fundamental to any system that aspires to understand the world the way humans do.
That thesis has attracted substantial capital. World Labs raised $230 million at its founding and closed a follow-on round of $1 billion in February 2026, bringing total funding to $1.23 billion. The investors backing that trajectory clearly believe that spatial intelligence — the ability to reason about objects, their positions, their physics, and their interactions — is not a niche capability but a foundational requirement for the next generation of AI.
Li has been explicit that large language models, however impressive their verbal reasoning, are fundamentally impoverished when it comes to physical understanding. Text corpora cannot teach a model what it feels like for a ball to roll off a table, or how light behaves when it passes through a glass prism, or how a robot arm must orient itself to pick up an irregular object from a cluttered surface. Atlas is the company’s answer to that gap.
Practical Applications
The most immediate use cases for Atlas span several high-value verticals.
In robotics and autonomous systems, a model with native 3D understanding could dramatically accelerate the training and deployment of physical robots. Current robotics pipelines rely on painstaking simulation environments or expensive real-world data collection. A world model like Atlas can generate photorealistic, physically consistent scenarios at scale, providing training data that is both cheap and diverse.
In film and media production, the ability to generate 1440p video that maintains 3D consistency opens up new possibilities for visual effects, virtual production, and synthetic media creation. Studios that currently spend millions on CGI for spatial environments could augment those workflows with generated content indistinguishable from real footage.
In autonomous driving, world models have already emerged as a key architectural component in training data pipelines. Atlas’s combination of video generation quality and 3D spatial understanding positions it well for partnerships with automotive companies seeking to test edge cases without putting vehicles on real roads.
In architecture and urban planning, the ability to generate and manipulate 3D environments from textual or visual prompts could transform how architects visualize designs and how urban planners model density, traffic flow, and environmental impact.
The Competitive Landscape
Atlas enters a market that is rapidly becoming crowded. OpenAI, Google DeepMind, and several well-funded startups are all working on world models and video generation systems. OpenAI’s GPT-6 Astra, released just days before Atlas, demonstrated strong multimodal capabilities. Google’s Gemini 3.8 Flash, also released in early September, added new real-time capabilities to its existing spatial understanding features.
What distinguishes World Labs is its singular focus. Unlike OpenAI or Google, which are building general-purpose AI platforms across dozens of application domains, World Labs has staked its entire existence on spatial intelligence. That focus, Li has argued, allows the company to go deeper on the hard problems of 3D understanding than any competitor building a generalist system can afford to.
Whether that thesis proves correct will depend partly on technical execution and partly on whether the market rewards spatial specialization or gravitates toward integrated platforms from larger players. The $1.23 billion in funding buys World Labs time to find out.
What Comes Next
Atlas is currently available in early access to select partners and developers. World Labs has not announced a general availability timeline or pricing, though the company has signaled that it is targeting both enterprise API customers and developers building spatial applications on top of the Atlas platform.
The launch positions World Labs as one of the most serious challengers in the emerging field of physical AI — systems designed not just to process language and images but to understand and interact with the world as a physical place. That is, ultimately, the kind of intelligence that robots need, that autonomous vehicles need, and that any AI system aspiring to operate in the real world rather than a text buffer will eventually require.
Fei-Fei Li has spent her career making the case that vision is not an afterthought in artificial intelligence. With Atlas, that argument has a compelling new exhibit.
*Word count: approximately 1,020 words*
