AI Model Intelligence API
Versioned model metadata, developer-reported evaluations and original TNL model reports. Sources currently cover curated public developer repositories on Hugging Face. This is not a complete Hub catalog or an independent benchmark service.
All /v1/intel/models routes use your existing TNL Bearer API key, verified-email requirements and plan quota. IDs are stable hf-… identifiers returned by the collection, not repository names containing slashes.
| Route | Returns |
|---|---|
GET /v1/intel/models | Latest collected revision per model |
GET /v1/intel/models/:id | Metadata, source commit, AI/ADES status and exact paper references |
GET /v1/intel/models/:id/versions | Retained source revisions |
GET /v1/intel/models/:id/evaluations | Structured model-card evaluation observations across revisions |
GET /v1/intel/models/:id/reports | Original published reports |
curl 'https://theneuralledger.com/v1/intel/models?publisher=Qwen&limit=10' \
-H "Authorization: Bearer $TNL_API_KEY"Collection filters: publisher, task, q, since, before, limit (1–100), and cursor. Responses include data, nextCursor and schemaVersion. Pass the opaque returned cursor for older observations. Dates filter repository update time for models/versions, observation time for evaluation rows, and publication time for reports. Evaluation rows can be filtered by q (dataset) and dates; publisher/task filters are not accepted there. Invalid filters return 400, missing/invalid authorization 401, unknown model 404 and temporary storage failure 503.
Model fields include repository/publisher identity, exact source commit, created/updated/observed timestamps, task, declared license, stored parameter count when reported by the source, base-model references, and linked arXiv identifiers. Missing values remain null. Repository timestamps are not independently established release dates. A stored tensor parameter count is not an independently verified active-parameter count or hardware requirement.
Evaluation rows retain task, dataset, metric, reported value/unit, source revision, observation time and available protocol conditions. verification: developer_reported means the source author supplied the claim. Neither Hugging Face hosting nor TNL digestion independently validates it. Numeric-looking strings are retained as strings, not guessed as normalized measurements. Missing units remain null.
comparisonEligible requires complete protocol metadata; comparisons additionally require identical task, dataset, metric, unit and conditions. Identical benchmark names alone never justify a ranking. This release does not compute leaderboards, performance deltas or run inference tests. Unsupported hardware, VRAM and speed measurements remain unknown.
Each model revision carries intelligence with pending/processing/retry/complete coverage, source evidence references, resolved or unresolved entities, affected assets where supported, potential effects and limitations. Shared detail is available at /v1/intel/digests/:key, with collection filter kind=ai_model. Raw model-card bodies and internal source snapshots are withheld.
Original reports appear at AI Model Intelligence. Publication requires completed AI/ADES digestion, sufficient evidence and a separate grounding review. Initial coverage and later source updates are labeled explicitly. Earlier reports and source revisions remain permanently available. Model-card links to papers are preserved; collected exact paper matches also expose a research API identifier.