The AI Model Atlas is a free, interactive graph that plots tens of thousands of Hugging Face models as connected nodes, showing which ones were fine-tuned, quantized, or adapted from which others. It comes out of a Hebrew University of Jerusalem research project, and despite what the "3D graph" phrasing in searches for this topic implies, the public demo actually renders as a 2D network you pan and click through in a browser — not a three-dimensional space.
Short answer: The AI Model Atlas is a free research demo (at huggingface.co/spaces/Eliahu/Model-Atlas) that graphs relationships between published machine learning models — which ones were fine-tuned, quantized, or merged from which. It's built from an initial dataset of over 60,000 models and 65,000 edges, rendered with Sigma.js and Gephi as a 2D network graph, not a literal 3D visualization.
In my testing, I opened the demo, searched for a handful of models I actually recognize, and clicked through the resulting clusters to see what the tool gets right and where the "3D" framing floating around this topic doesn't match what you'll actually see on screen. Below is what to expect before you start, the steps I used to explore it, and where I think it's actually worth your time.
What you'll need
You don't need an account, an API key, or anything installed — the AI Model Atlas demo runs entirely in the browser once the page loads. What helps:
- A modern browser and a few extra seconds of patience on first load. The canvas is rendering tens of thousands of nodes with Sigma.js, so it's noticeably slower than a typical webpage.
- One or two model names you actually care about. Searching blind is a lot less useful than starting from something you'd consider using, like a model you've already deployed or one you're comparing against alternatives on my best AI models page.
- Optional: Gephi (free, desktop) if you'd rather open the underlying GraphML export yourself than use the hosted demo.
- Optional: an AI research assistant to help interpret what a cluster of fine-tunes actually represents — I cover that workflow in the prompts section below.
Step-by-step: exploring the AI Model Atlas
1. Open the demo
Go to the Model Atlas Space on Hugging Face. It's a public Space built by researcher Eliahu Horwitz, free, with no sign-in required.
2. Let the full network render
The default view loads a large connected network. Give it a moment — I timed my own first load at well over ten seconds on a normal broadband connection, since the page is drawing tens of thousands of nodes onto a single canvas before it becomes interactive.
3. Search by name instead of browsing
Use the "Search by name" box rather than trying to visually scan the whole graph — with this many nodes packed together, scanning by eye just isn't productive. I searched a few widely-used models (bert-base-uncased, gpt2, distilbert-base-uncased) and each one snapped the view to its local neighborhood instead of the full network.
4. Read the Information Pane
Clicking a node opens an Information Pane with that model's details and its direct connections. This is where the graph earns its keep: you can see at a glance which checkpoints a model descended from, rather than digging through a model card's text.
5. Change what determines node color and grouping
The "Color By" and "Group Selector" dropdowns let you re-render the same graph by different attributes instead of hunting for a fixed legend. Switching Color By is the fastest way to spot a pattern — I found it more useful for noticing dense fine-tune clusters than the default view.
6. Reset and try the "Select Atlas" dropdown
The demo isn't one single graph — a "Select Atlas" dropdown switches between different connected components the researchers charted. If a search dead-ends or the view gets cluttered, "Return to the full network" resets you to the starting point.
Searches and prompts you can copy
These are exact searches and follow-up prompts close to what I used while testing:
- Search
bert-base-uncasedin the Model Atlas search box — it sits at the center of one of the largest fine-tune clusters, so it's a good first stop to see how the graph reads. - Search
gpt2and compare its neighborhood size todistilbert-base-uncased— the difference in cluster density is a quick, visual way to see which base model spawned more public derivatives. - Search
stabilityai/stable-diffusion-xl-base-1.0if you want to see how the graph handles an image model instead of a text model. - Once you've found a cluster, copy the model IDs you see into an AI assistant with a prompt like: "I searched '[model name]' in the Hugging Face Model Atlas and it shows several connected nodes labeled as fine-tunes. Explain what a fine-tune edge means here and why someone would pick one of these derivatives over the base model." My how to use Perplexity for research guide covers getting cited, sourced answers back for exactly this kind of follow-up.
- For a side-by-side on two specific model IDs from the same cluster, try: "Compare [model A] and [model B] from Hugging Face — what's the practical difference for someone choosing between them for [your use case]?"
AI Model Atlas vs. other model-visualization tools
The "3D" search intent around this topic usually means one of a few different things, so here's how the Model Atlas demo actually compares to the tools people mean when they ask for one:
| Tool | Maker | What it graphs | Visualization | Cost |
|---|---|---|---|---|
| AI Model Atlas (demo) | Eliahu Horwitz et al., Hebrew University of Jerusalem | Relationships between 60,000+ published models (fine-tunes, quantizations, merges) | 2D network graph (Sigma.js + Gephi export) | Free |
| Google Model Explorer | Google AI Edge | Internal layer/op graph of one model (JAX, PyTorch, TensorFlow, TFLite) | Hierarchical graph, WebGL-accelerated | Free |
| Zetane Viewer | Zetane Systems | Internal architecture and tensors of one model (ONNX and others) | 2D and 3D | Free desktop app |
| LLM Visualization (bbycroft.net) | Brendan Bycroft | Step-by-step internals of a single GPT-style model | 3D animated walkthrough | Free, browser-based |
The real distinction isn't 2D versus 3D — it's population-level versus single-model. The Model Atlas is the only one of these that maps how thousands of separate models relate to each other; the other three open up the inside of one model at a time. Of the four, bbycroft's LLM Visualization is the one that's actually rendered in 3D. If you're deciding which model to run rather than how one is built, that population-level view is the more useful of the two questions to ask, and my Flint writeup covers a separate free tool for turning whatever you find into your own chart.
Common mistakes to avoid
The first thing I got wrong was expecting a literal 3D space to fly through — it's a 2D canvas rendered with Sigma.js, and no amount of scrolling turns it into one. Second, don't expect a finished, polished product: the project's GitHub repository describes its released code as "an initial rough version of the preprocessing code," with the full interactive atlas still coming to the codebase separately from the hosted demo. Third, if a model you search for doesn't show up, that's not necessarily a bug — the underlying paper notes that most of the wider model population remains undocumented, so the charted portion is a fraction of what's actually hosted. Fourth, treat the dataset as a snapshot, not a live feed: it covers an initial set of roughly 60,000 models and 65,000 edges the researchers charted, not the full, constantly growing Hugging Face catalog in real time. Fifth, don't read node size as a quality signal — it reflects download counts and graph structure, not benchmark performance, so a big node isn't automatically the model you should use.
Tools that make this easier
If tracing a model's lineage in the Atlas leaves you deciding what to actually deploy, cross-check it against a hands-on ranking instead of the graph alone — my best AI models guide tests and ranks the models people search for most, independent of how many fine-tunes they've spawned. If your interest in Hugging Face runs toward what's inside these repositories rather than how they connect to each other, I wrote a separate piece on scanning Hugging Face datasets for exposed secrets that covers a different kind of audit of the same platform. For turning what you find into an actual report or chart rather than a screenshot, Flint is a free, open-source tool worth having open alongside the demo. And if you want an AI assistant to help you dig into an unfamiliar model's documentation while you've got a node's Information Pane open, see how to use Perplexity for research or how to use Google AI Studio if you'd rather run that comparison against a Gemini model instead — my AI tool reviews page explains how I test tools like these before recommending them.
My take
The AI Model Atlas is a genuinely useful research tool for a narrow question — how did this specific model come to exist, and what else got built from the same starting point — and it answers that question faster than reading through model cards one at a time. It is not, despite how the topic gets searched, a 3D visualization, and it's not a finished consumer product either; the researchers themselves describe the released code as an early version. If you work with Hugging Face models regularly and want to understand lineage at a glance, it's worth bookmarking. If you were hoping for a slick, navigable 3D universe of AI models, the closest thing on this list is bbycroft's LLM Visualization, and it visualizes one model's internals, not a population of them.
Frequently Asked Questions
Is the AI Model Atlas free to use?
Yes. The hosted demo on Hugging Face Spaces is free and public, with no account or payment required to search it.
Is the AI Model Atlas graph actually 3D?
No. The public demo is built with Sigma.js and renders as a 2D network graph on a canvas element, based on a Gephi export. If you specifically want a 3D model visualization, bbycroft's LLM Visualization is the one on this list that actually renders in three dimensions.
How long does it take to explore one model's lineage?
A few minutes. Once the full network finishes its initial load, searching a specific model name and reading its Information Pane takes well under a minute — the slow part is the first page load, not the search itself.
What's the easiest way to try it without installing anything?
Open the Hugging Face Space directly in your browser. Everything runs client-side; you don't need Gephi, Python, or any local setup unless you want to work with the underlying GraphML data yourself.
Can I get the underlying dataset myself?
The researchers have released datasets and preprocessing code on GitHub and Hugging Face, though the repository's own README describes the current code as an early, rough version rather than a polished package — expect to do some assembly if you want to rebuild the atlas locally instead of using the hosted demo.