Model Catalog

A vendor-neutral, kind-aware public catalog of LLM / embedding / rerank / media models — which model ids exist per vendor and, for each, its kind (chat, embedding, rerank, image, transcription, speech, video, moderation), context window, max output tokens, embedding dimensions, modalities and capability hints. Published as a free, unauthenticated, CORS-open, versioned JSON artifact. Part of Viglet Turing ES.

Identity + kind, not pricing. The catalog never carries cost/price — see docs/STRATEGY.md §I.

Public API

URL	Meaning
https://openviglet.github.io/model-catalog/catalog.json	Rolling latest
https://openviglet.github.io/model-catalog/catalog-v1.json	Pinned to schema v1
https://openviglet.github.io/model-catalog/catalog.schema.json	JSON Schema (Draft 2020-12)
https://openviglet.github.io/model-catalog/	Browsable reference page

Full contract → docs/references/api.md.

# every embedding model across all vendors
curl -s https://openviglet.github.io/model-catalog/catalog-v1.json \
  | jq '.vendors | to_entries[].value[] | select(.kind=="EMBEDDING") | .id'

Client SDKs

Zero-/minimal-dependency, read-only clients wrap the endpoints above (URL selection, vendors-map flattening, byKind/byVendor/get filters, in-memory caching). All Apache-2.0, published to their public registries — nothing below needs auth:

SDK	Install
Java (JDK HttpClient, 17+)	Maven Central — io.github.openviglet:model-catalog-client
JS / TS (ESM, Node 18+ & browser)	npm — npm i @openviglet/model-catalog-client
Python (stdlib urllib, 3.8+)	PyPI — pip install openviglet-model-catalog-client

<dependency>
  <groupId>io.github.openviglet</groupId>
  <artifactId>model-catalog-client</artifactId>
  <version>1.0.2</version>
</dependency>

Per-SDK docs: Java · JS/TS · Python.

Layout

catalog/     canonical model-catalog.json + JSON Schema  (single source of truth)
pipeline/    zero-dep multi-source regeneration pipeline  (propose-and-review)
scripts/     emit.mjs — builds the public artifacts from the canonical file
public/      index.html (browsable page); models/* is emitted (gitignored)
docs/        references (api, pipeline) + roadmap set (ROADMAP/CHANGELOG/…)

Maintaining the catalog

The canonical file is hand-editable for a quick fix, or regenerated from multiple trustworthy sources per vendor (live vendor APIs + the LiteLLM registry + curated overrides.json) via the propose-and-review pipeline:

npm run regen              # dry-run: fetch → merge → validate → write report (canonical untouched)
npm run regen -- --apply   # write the canonical file (the only path that overwrites it)
npm run emit               # rebuild the public artifacts from the canonical file
npm test                   # pipeline unit tests

A missing API key skips that source; nothing hard-fails. The canonical file is only ever overwritten on --apply. Full workflow → docs/references/pipeline.md.

Publishing

.github/workflows/publish.yml emits the artifacts and deploys public/ to GitHub Pages on every push to main that touches the catalog. It never regenerates from upstream — publishing only serves the already-reviewed canonical file.

Zero dependencies

The pipeline, emit step and browsable page use only Node built-ins + global fetch and inline HTML/CSS/JS. There is nothing to npm install. Requires Node ≥ 20.

Community & governance

• CONTRIBUTING.md — how to propose or correct a model (no code required) and how a proposal becomes a catalog entry.
• GOVERNANCE.md — who maintains this community-owned reference and how decisions are made (the neutrality guarantees).
• CODE_OF_CONDUCT.md — Contributor Covenant.
• SECURITY.md — how to report a vulnerability privately.

License

• Code — Apache-2.0 (the pipeline, emit step, page and SDKs).
• Catalog data — CC-BY 4.0 (the JSON compilation and every derived artifact): reuse freely with attribution to the Viglet Model Catalog.

The data is provided as is; accuracy is not warranted. Figures are indicative and cited — verify against the vendor before relying on any value (see each entry's sources / lastVerified).