ทันเอไอ

Researchers Unveil AI Model Atlas, a 3D Map of Over 100,000 AI Models on Hugging Face

A research team from The Hebrew University of Jerusalem has built an interactive visualization tool that reveals the 'family trees' of more than 100,000 AI models in an explorable 3D graph accessible via browser

The world of AI now sees hundreds of thousands of new open-source models emerging, to the point where almost no one can keep track of which model came from where. A research team led by Eliahu Horwitz from The Hebrew University of Jerusalem has launched the AI Model Atlas project, an interactive visualization tool that compiles a massive number of Machine Learning models — including over 100,000 NLP (Natural Language Processing) models on the Hugging Face platform — and displays them as an interconnected 3D graph.

This map does more than just plot model points to look pretty. It also shows the 'lineage' of each model: which model was fine-tuned from which, which models were created by merging multiple models together, and which architectures were inherited from which predecessors. Users can click, zoom, and explore model clusters directly through the browser, powered by the Cosmograph/WebGL 3D graph technology.

Behind the project is the concept of Weight-Space Learning — learning by exploring the 'weights' of existing models rather than training new ones from scratch. The team believes this will promote Model Recycling, the practice of building new models on top of old ones, reducing both the cost and energy wasted on redundant model training.

The project has clear academic credentials, with a research paper published on arXiv (No. 2503.10633) and a presentation at NeurIPS, a leading international AI conference, under the concept 'We Should Chart an Atlas of All the World's Models' — an effort to push mapping the world's AI models into becoming an industry standard.