Hugging Face Public Model Revision Evidence API

Detect declared model-repository revision and metadata changes without downloading model weights or paying for inference. Returns the Hub commit SHA, last-modified value, gating flag, pipeline/library tags, selected architecture declarations and license tags. Useful for model inventory QA and tracking known public repositories in a deployment review.

One GET huggingface.co/api/models/owner/model per distinct ID, at most5. Response identity, public flag and revision shape are checked. Public metadata for a gated model may be visible even when its weights require access: a returned row is not permission to download or use those weights. No token, access request, gating bypass, inference, safety/compatibility testing or license verification. Architecture/license fields are declarations; missing fields remain empty/unknown. At most200 tags and100 architectures; tag truncation makes complete=false, excessive architectures reject that source. Download counters, uploader/author identity, card body, training data, discussion and file tree are omitted. Model rename redirects across hosts are refused, and response aliases with a different identity are rejected.

Quick start

{"modelIds": ["sentence-transformers/all-MiniLM-L6-v2", "google-bert/bert-base-uncased"]}

Pricing and budget

0.001 USD per emitted public model observation (1 USD per1000), plus0.005 USD startup at default256MB. Higher memory can multiply startup events; check effective Store pricing. Failed/empty sources do not emit result charges; startup can still apply. At zero result budget, input validation occurs but no upstream request is sent, and rowsAvailable is null (unknown). A positive budget can request the bounded source set before output; emission stops at the result cap and complete=false/PARTIAL_BUDGET reports truncation.

Snapshot, integration and limits

Use dataset rows together with OUTPUT and SNAPSHOT. Pass SNAPSHOT.previousHashes as previousHashes in a later run. NO_BASELINE/FIRST_SEEN/UNCHANGED/CHANGED compare only selected metadata including itemKey and sourceUrl; missing records are not deletions. FIRST_SEEN means absent from the supplied baseline. There is no persistent hidden state or automatic polling. Export JSON/CSV/Excel through the default dataset.

API example: POST https://api.apify.com/v2/actors/soilair~hugging-face-public-model-revision-evidence-api/runs with your Apify Authorization header and the JSON input above. Poll the returned run; read dataset items and OUTPUT/SNAPSHOT records using its store IDs. Do not place tokens or confidential data in input. No upstream key is needed. Results live in Apify run storage according to retention; no separate analytics service.

Only fixed HTTPS provider endpoints are constructed from validated identifiers. Public-IP DNS validation, pinned connections, TLS hostname verification,2MiB decoded responses, approximately95 seconds of network budget, at most5 same-provider redirects. Official API calls are distinct from website crawling; no arbitrary URLs or HTML pages. No linked documents/files are fetched. Queries are sequential with1-second spacing. Provider403/429 or Retry-After stops later calls; retry instructions are reported and no retry/proxy/circumvention is attempted. Shared IP limits and changes to upstream services can still block runs. Caller must respect provider headers and cache/reduce repeated requests. Text is data and should not be executed.

Developed with AI assistance. Tests and live/cloud comparisons qualify the selected-field contract; they do not prove commercial demand, provider uptime or a legal/license determination. Support: share a run ID and expected fields, without credentials or private material.

References: https://huggingface.co/docs/hub/api · https://huggingface.co/.well-known/openapi.json