Hugging Face Model Monitor - Hub SHA Watchlist

Re-check the Hugging Face Hub models you already care about. Get SHA, downloads, license and visibility changes — and skip the rows that did not move.

This Actor takes a caller-owned list of Hub model IDs, reads the public Hub API, and writes one structured row per model. Pass the previous run's snapshot back in and you get INITIAL, UNCHANGED or CHANGED on every successful check. No Hub token. No weight downloads. No Hub crawl.

---

What this Actor does

You already know the model IDs (gpt2, owner/name). You want a scheduled check that tells you when the commit SHA, license, gated flag, download count or visibility moved — without paying a start fee every morning and without pulling gigabytes of weights.

That is the job: a watchlist monitor, not a Hub dump scraper.

---

What you get

{
  "schemaVersion": 1,
  "id": "gpt2",
  "status": "ok",
  "modelId": "gpt2",
  "pipelineTag": "text-generation",
  "libraryName": "transformers",
  "downloads": 18423011,
  "likes": 2144,
  "sha": "607a30d783dfa663caf39e06633721c8d4cfcd7e",
  "lastModified": "2026-04-17T17:44:44.000Z",
  "gated": false,
  "disabled": false,
  "license": "mit",
  "author": "openai-community",
  "changeType": "INITIAL",
  "changedFields": [],
  "checkedAt": "2026-09-06T12:00:00.000Z",
  "sourceUrl": "https://huggingface.co/api/models/gpt2",
  "error": null
}

gpt2 is a public example identifier, not a customer or endorsement. Missing source strings are null — never "N/A".

---

Main use cases

• Model-card drift. Catch SHA and license changes on models your product actually loads.
• Gated / disabled alerts. See when a watched model becomes gated or disabled.
• Download and like movement. Treat Hub counters as a change signal, not a ranking.
• Scheduled MLOps. Put the Actor on your own cadence and keep the snapshot in your workflow state.
• n8n / MCP / backend jobs. Start the run from a server that holds APIFY_TOKEN.

---

Quick start

{
  "models": ["gpt2"],
  "maxItems": 1
}

Run it, open the dataset, and save NEXT_SNAPSHOT from the default key-value store.

---

Repeat workflow

1. Keep one Hub model watchlist per customer or project (1–50 IDs).
2. Start a bounded run. Schedule it yourself if you need a cadence.
3. Save the returned NEXT_SNAPSHOT JSON in your private state.
4. On the next run pass it as previousSnapshot and set onlyChanges: true.
5. Route changeType: CHANGED rows (SHA, visibility, downloads, license) for review.

No schedules are enabled by default. Snapshots use the run's default storage, never a shared named store. A repeated check without a supplied snapshot is a new initial check.

---

Input

Field	Type	Default	What it does
models	array	—	1–50 Hub IDs (name or owner/name). Duplicates checked once. Letters, digits, ., _, - only.
maxItems	integer	50	1–50. Must cover all unique IDs; a smaller value is rejected rather than silently dropping IDs.
onlyChanges	boolean	false	Suppress unchanged successful records without charging them. The first observation is still emitted.
previousSnapshot	object	—	Prior NEXT_SNAPSHOT with schemaVersion: 1, kind: "hf-model", and a records map.

---

Pricing

Event	Price	When it happens
Successful watchlist result	$0.003	One successful model check written to the dataset
Actor start	none	No custom start fee

You are not charged for: failed requests, parse failures, retries, duplicate IDs, invalid input, or unchanged records suppressed by onlyChanges. With onlyChanges: false, an unchanged successful refresh is delivered work and is charged. The run stops when eventChargeLimitReached is true.

Worked example. 50 models every weekday, onlyChanges: true after the first run, 4 changed rows/day → about (4 × $0.003) = $0.012 per run, roughly $0.26/month in Actor event charges if that pattern holds.

How this compares

Actor	Price per result	Per-run start	Weights downloaded
This Actor	$0.003	none	No
parseforge/huggingface-model-scraper (Store card, 2026-09-06)	$0.005	$0.005	Dump-oriented listing

Competitor list prices change. Check the current Store cards before relying on this table. This Actor is a caller-owned watchlist, not a Hub catalogue dump.

---

Using the API

Keep APIFY_TOKEN on a protected backend. Set maxItems and maxTotalChargeUsd.

curl --fail -X POST "https://api.apify.com/v2/acts/technicaldost~huggingface-model-watchlist-monitor/run-sync-get-dataset-items?token=$APIFY_TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"models":["gpt2"],"maxItems":1}'

JavaScript

import { ApifyClient } from 'apify-client';

const client = new ApifyClient({ token: process.env.APIFY_TOKEN });
const run = await client.actor('technicaldost/huggingface-model-watchlist-monitor').call({
    models: ['gpt2'],
    maxItems: 1,
});
const { items } = await client.dataset(run.defaultDatasetId).listItems();

Python

from apify_client import ApifyClient
import os

client = ApifyClient(os.environ["APIFY_TOKEN"])
run = client.actor("technicaldost/huggingface-model-watchlist-monitor").call(run_input={
    "models": ["gpt2"],
    "maxItems": 1,
})
for row in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(row["modelId"], row["sha"], row["changeType"])

n8n. HTTP Request → POST https://api.apify.com/v2/acts/technicaldost~huggingface-model-watchlist-monitor/runs with header Authorization: Bearer {{$credentials.apifyToken}}. Poll the run, then read the default dataset and NEXT_SNAPSHOT.

MCP / agents. Call the Actor asynchronously from a server tool. Do not put the token in the browser or in a public agent transcript.

---

Limitations

• Public Hub metadata only. Private repositories are written as errors, not guessed.
• Weights, datasets, Spaces and files are out of scope.
• Hub availability and freshness govern results. This is not a ranking, safety, or license-compliance opinion.
• One request at a time, 20 seconds per request, 1.5 MB per response, no automatic retries.
• At most 1,000 snapshot records are retained; SUMMARY.snapshotPruned reports eviction.

Not affiliated with or endorsed by Hugging Face.

---

Related Actors

• Software End-of-Life Monitor — support-window changes on products you already run.
• crt.sh Certificate Monitor — new public TLS certificates for hostnames you already own.

---

Changelog

See CHANGELOG.md. Release 1.0.0 (2026-09-06): caller-owned Hub watchlists, SHA/visibility deltas, isolated snapshots, $0.003 per successful result, no start fee.