The Atlas model represents a multi-modal world model built from scratch to establish a new standard in the construction and generation of 3D spatial environments.
Article Index:
- Launch of the Atlas model and the horizon of spatial intelligence
- Unified computational architecture and 3D generation
- Spatial reconstruction and outperforming competitors
- World Labs’ evolution and billion-dollar funding rounds
- Simulation and robotics applications alongside existing caveats
- Frequently asked questions
Launch of the Atlas model and the horizon of spatial intelligence
On Tuesday, artificial intelligence startup World Labs unveiled its innovative model “Atlas”, a multi-modal AI system specifically designed to generate, reconstruct, and simulate 3D spatial environments with unprecedented geometric precision, in a move the company described as opening new and crucial horizons in what has become known as “spatial intelligence”.
Dr. Fei-Fei Li, leading AI scientist and co-founder of the company, announced the rollout of the new model via her account on the X platform, describing it as the first multi-modal world model built and trained entirely from scratch. She proudly asserted that it represents the best computer model ever to control virtual camera movement in 3D space, paving the way for a comprehensive visual revolution.
Unified computational architecture and 3D generation
World Labs explains that Atlas relies on an innovative architecture known as “multi-modal auto-regressive diffusion transformers”, a unified software framework capable of processing text inputs, still images, video footage, and 3D spatial data within a cohesive and shared spatial coordinate framework.
The model can generate animated cinematic video clips at a stunning resolution of up to 1440p for a duration of up to a full minute, relying on anywhere from just 1 to 6 reference images along with a user-defined virtual camera path with high geometric precision. This grants creators the freedom to direct scenes and navigate through fictional worlds with absolute seamlessness and the highest standards of realism.
Spatial reconstruction and outperforming competitors
Atlas’s capabilities extend beyond creative generation to include the precise spatial reconstruction of real-world scenes using very scarce and limited inputs. The model possesses the ability to reconstruct and render an entire environmental scene based on just two or three images, producing entirely new viewing angles that were not visible in the original footage, alongside advanced depth maps, precise point clouds, and 3D volumetric representations.
The company published benchmarking test results confirming Atlas’s superiority over specialized open-source models across globally accredited evaluation suites. In independent human evaluation tests for camera-guided generation, judges preferred Atlas by wide margins compared to major market competitors, including the MiniMax H3, Gemini Omni Flash, and SVDance 2.5 models.
World Labs’ evolution and billion-dollar funding rounds
Atlas represents a natural extension and development of the company’s first commercial model, “Marble”, which debuted publicly in November 2025 to enable the construction of 3D worlds across multiple media, followed by the launch of a public API in January 2026 for creating freely explorable virtual environments.
In July 2026, World Labs strengthened its technological foundation by fully acquiring the company “Cenex” to enhance its capabilities in robotics-oriented spatial intelligence. The startup successfully raised massive investment funding totaling $1.23 billion, which included a massive $1 billion funding round in February 2026 featuring a $200 million contribution from Autodesk, with active participation from chip industry giants AMD and Nvidia, and the Fidelity financial institution.
Simulation and robotics applications alongside existing caveats
World Labs aims to deploy Atlas in the fields of filmmaking and visual effects, 3D video game development, industrial and architectural design, and advanced robotics engineering. The model can generate sensor readings and visual feedback that simulate a robot’s movement path in reality, contributing to the training of autonomous systems via safe simulation environments.
Despite this enthusiasm, Atlas’s performance claims have not been subjected to independent testing and auditing by neutral parties, nor has the company disclosed commercial pricing plans, computational latency, or the computing power requirements necessary for operation, noting that the model will initially be made available to selected partners before being integrated into future releases.
Frequently asked questions
Q: What is the Atlas model and what are its primary use cases?
A: It is a multi-modal spatial intelligence model that accurately generates and simulates 3D environments to develop visual effects, games, and robotic systems.
Q: How does Atlas handle sparse inputs to reconstruct scenes?
A: Atlas can reshape complete realistic environments based on just two or three images while extracting depth maps and spatial representations.
Q: What is the total funding World Labs has raised so far?
A: The company has raised approximately $1.23 billion with support from major firms such as Autodesk, Nvidia, AMD, and leading investment funds.
