Focusing on spatial intelligence! "AI Mother" Fei-Fei Li releases the first commercial world model

"AI Goddess" Fei-Fei Li's startup has taken a key step in commercialization on the "world model" track. Her company, World Labs, has officially launched its first commercial product, Marble, marking the expansion of artificial intelligence from abstract language text understanding to the realm of spatial intelligence that simulates and interacts with the physical world, aiming to gain an edge in the next generation of AI technology competition.

On November 12th, World Labs, co-founded by Stanford University professor Fei-Fei Li, announced that its first product driven by a multimodal world model, Marble, is officially open to the public. This model can utilize various input forms such as text, photos, videos, or 3D layouts to generate editable and downloadable 3D interactive environments, giving World Labs a first-mover advantage in competition with tech giants like Google.

Compared to the limited preview version released two months ago, the official version of Marble has achieved significant functional expansions, including support for larger-scale multimodal inputs and the launch of a creative center called Marble Labs. In terms of business model, the product adopts a combination of freemium and paid subscription, aiming to quickly enter the market and establish a user base.

The launch of Marble directly targets industries with a huge demand for 3D content, such as game development, visual effects (VFX), and virtual reality (VR). It provides creators with a brand new asset generation tool, but also sparks in-depth discussions in the industry about how AI will reshape creative workflows and its potential impact on jobs.

Early Commercialization, Establishing Differentiation in Competition

In the burgeoning world model track, the commercialization release of Marble allows it to stand out from numerous research projects and technology demonstrations.

Currently, Google's Genie model is still in a limited research preview phase, while products from startups like Decart and Odyssey remain in free demo versions. According to Bloomberg, Marble is the first product in this track to be put into commercial application. Its core differentiation lies in its ability to generate persistent, downloadable 3D environments, rather than dynamically generating worlds as in World Labs' own real-time model RTFM (Real-Time Foundation Model) during user exploration.

The company stated that this design significantly reduces scene deformation and inconsistencies, and allows users to export results in professional formats such as Gaussian splats, meshes, or videos for integration into existing workflows.

Marble's input-output pipeline image source: World Labs

Focus on Creative Control and Multimodal Input

To address the pain point of "uncontrollable" AI-generated content, Marble has made "creative control" the core of its product. Justin Johnson, co-founder of World Labs, stated in an interview with TechCrunch: "We don't want machines to take over completely and deprive creators of their creativity."

To this end, Marble offers high flexibility at both the input and editing ends.

• Input: The official version supports users to upload multiple images or short videos to showcase a space from different angles, thereby generating a more realistic digital twin. This addresses the limitation of the preview version, which only supports single image input, requiring the model to "imagine" a lot of details on its own.
• Editing: The product has launched an experimental 3D editor called "Chisel." This feature allows users to first build a rough spatial structure (such as walls and blocks) and then fill in the visual style through text prompts, achieving a separation of structure and style. Johnson explained: "I can enter the scene, directly grab a 3D block representing a sofa, and move it."

Additionally, users can "expand" the generated world or stitch multiple independent worlds into a larger space under "Composer Mode," further enhancing creative freedom.

A spaceship environment created using Marble, with text prompts overlaid (note how the lighting realistically reflects on the hub walls).

Image source: World Labs/TechCrunch

Clear Business Model Targeting Three Major Application Scenarios

Marble's business path is clear, meeting different user needs through tiered subscriptions. Its subscription plans are divided into four tiers: the free version offers 4 generations per month; the standard version ($20/month) and professional version ($35/month) provide more generations and advanced features; while the flagship version ($95/month) unlocks all features and includes 75 generations and commercial usage rights.

In the short term, World Labs has locked its target market in three major areas:

• Game Development: Johnson believes developers can use Marble to generate background environments and atmospheric spaces, then import these assets into game engines like Unity or Unreal for further development, serving as a supplement rather than a replacement for existing pipelines.
• Film and Television Special Effects: Compared to AI video generators, the 3D assets generated by Marble provide artists with precise scene control and camera movement capabilities, avoiding inconsistencies in visuals.
• Virtual Reality (VR): Although VR is not the current primary focus, Johnson pointed out that the industry has an "extreme demand" for high-quality content. All worlds generated by Marble are compatible with mainstream headsets like Vision Pro and Quest 3 However, the market's acceptance of this technology is not entirely optimistic. According to a recent survey by the Game Developers Conference (GDC), one-third of respondents believe that generative AI has a negative impact on the gaming industry, with concerns mainly focused on intellectual property and declining content quality.

Fei-Fei Li's Ultimate Vision: Moving Towards "Spatial Intelligence"

Behind the release of Marble is Fei-Fei Li's grand vision for the future development of AI—achieving "Spatial Intelligence."

In a recent article, Fei-Fei Li elaborated that current large language models excel at processing abstract knowledge but lack a true understanding of the physical world, akin to "walking in the dark." She believes that endowing machines with the ability to understand, navigate, and interact with the three-dimensional world, namely spatial intelligence, will be key to achieving true general artificial intelligence. The world model is the core technology on the path to this goal.

According to her vision, world models with spatial intelligence will transform multiple industries in phases:

• Short-term: Empowering the creative industries by providing powerful tools for creators in film, gaming, and architecture.
• Mid-term: Promoting the development of embodied intelligent robots, making them collaborative partners for humans through large-scale training in simulated environments.
• Long-term: Triggering revolutions in fields such as science, medicine, and education by enhancing the capabilities of human experts through simulated experiments and assisted diagnosis.

"Without spatial intelligence, our dream of truly intelligent machines will be incomplete," Fei-Fei Li wrote. This vision has received strong support from the capital markets. Since its establishment in 2024, World Labs has rapidly completed approximately $230 million in financing, with a valuation exceeding $1 billion, entering the ranks of "unicorns." Its investors include industry giants and top venture capital firms such as a16z, NVIDIA NVentures, AMD Ventures, and Intel Capital