Home

By reading this wiki, you will understand why model discovery is a navigation problem — and how ModelAtlas solves it with signed hierarchies and anchor-based similarity. No embeddings, no GPU at query time. Just integers and set intersections.

---

What ModelAtlas Does

HuggingFace Hub has a million models and roughly ten structured fields per model. The relational information that makes models findable — capabilities, lineage, efficiency profiles, architectural properties, domain specializations — is trapped in unstructured model cards, naming conventions, config files, and community knowledge.

ModelAtlas extracts that information and encodes it as positions in a structured semantic space. The result is a navigable network where queries are movements through space ("toward smaller and more code-focused"), not WHERE clauses ("parameter_count < 3B AND tags LIKE '%code%'").

The entire system is a SQLite file, a vocabulary of ~170 anchor labels, and signed integers. No GPU at query time. No vector store in the background. No running services.

---

Choose Your Path

I want to use it

Start with Getting Started for installation and your first query, then browse Query Examples for the kinds of questions ModelAtlas can answer that HuggingFace cannot.

I want to understand how it works

System Overview gives the 30-second architecture map. From there, dive into the Query Engine (how scoring works), the Extraction Pipeline (how data gets in), or the Data Model (what the database looks like).

I want to understand why it's designed this way

The Gap defines the problem: what's missing from HuggingFace and why it matters. Signed Hierarchies introduces the core model. Emergent Similarity explains why anchor overlap replaces embeddings. Navigation Geometry makes the case for movement through space over filter queries.

---

The Concept in One Diagram

User asks: "small code model with tool-calling that runs on a Mac"

LLM decomposes into:
  EFFICIENCY  → negative (small)
  CAPABILITY  → positive (code, tool-calling)
  COMPATIBILITY → positive (MLX/Apple Silicon)
  require: [code-generation]
  prefer:  [tool-calling, Apple-Silicon-native]

ModelAtlas does:
  1. SQL pre-filter on required anchors
  2. Bank alignment scoring (signed distance from zero)
  3. IDF-weighted anchor overlap
  4. Multiplicative combination → ranked results

No embeddings computed. No model loaded. Just arithmetic on integers
and set intersections on small lists.

The intelligence is in the interaction — the LLM decomposes the question, ModelAtlas does the math. Neither is smart alone, but the system is.

---

Current State

ModelAtlas is in active beta. The network contains 19,498 models with 166 anchors across 8 banks, 128K+ model-anchor links, and a multi-phase correction pipeline actively improving accuracy. See Data Distribution for how to get the pre-built network.

---

Glossary · Getting Started →