Hugging Face Trending Models Scraper — Radar & Change Watch

Turn a public Hugging Face model list into a readable discovery report and a repeatable metadata snapshot. Find models in trending, download, like, creation or modification order, then save a named watchlist to review what changed at the next successful run.

Useful for a research shortlist, a newsletter source list or a recurring model discovery review. Every live row links to its source model page. This community Actor is independent of Hugging Face.

Start with live data

Click Start with the default input. It collects up to 20 public models in Hugging Face trending order. No Hugging Face token, proxy, AI key or paid source Actor is needed.

{
  "maxResults": 20,
  "sort": "trendingScore",
  "demoMode": false
}

Open Visual model radar after the run, or use the default dataset, JSON report and CSV snapshot. The report includes source time, query, coverage warnings and comparison status.

Inputs

Input	Default	Meaning
maxResults	20	Up to 100 models from a bounded query.
sort	trendingScore	Descending trendingScore, downloads, likes, lastModified or createdAt.
task	blank	Optional exact Hub pipeline tag, such as text-generation.
search	blank	Optional server-side text search within model IDs.
reportName	AI model radar	Display name, 1–80 characters.
watchlistName	blank	Optional persistent comparison name, up to 80 characters.
demoMode	false	Five fictional offline rows; no live collection or saved history.

Task tags are limited to 80 characters and search to 100. Filters can return zero models. This version covers models only, with no separate author filter or exact-repository inventory mode.

Compare a named watchlist

Run the same input twice, using a stable watchlist name:

{
  "maxResults": 20,
  "sort": "trendingScore",
  "task": "text-generation",
  "watchlistName": "weekly-text-generation",
  "reportName": "Text generation radar"
}

The first live run establishes a baseline. A later comparable run records models newly present in the returned results, changes in position and metrics, and observed changes to revision SHA, declared license IDs, access setting, task or library. Unknown fields remain unknown. A missing declaration does not establish unchanged access or license facts.

History is stored in a named Apify key-value store. Keep one run at a time per watchlist. Changing sort, task, search or result limit starts a new baseline. Failed or invalid collections do not advance it; incomplete selections suppress history advancement. Demo runs never write history. Comparisons cover saved observations, not every intervening source change.

Outputs

Output	Contents
Default dataset	One normalized model observation per row.
REPORT_HTML	Standalone visual report with model links, metrics, change review and limitations.
REPORT_JSON	Rows, query, summary, warnings, history status and models absent from current results.
MODELS_CSV	Spreadsheet-friendly model snapshot with guarded formula-like text.

Rows include modelId, modelUrl, sourcePosition, task, library, licenseIds, gated, revisionSha, downloads30d, likes, trendingScore, createdAt, lastModified, observedAt, change, prior position, metric differences and metadataChanges. Missing metrics are null; a real zero stays zero. JSON preserves structured license and change arrays.

Read the signals correctly

• Position is rank within the current bounded query, not a global quality or performance ranking. The Actor preserves source order and does not invent a trending-score formula or time window.
• downloads30d is Hugging Face's rolling 30-day request-count metric. It is not installs or unique users. Its snapshot difference can be negative and is not newly acquired downloads. Download counting documentation.
• A revision SHA or modification time can change after a repository edit. It does not prove new weights or a model release.
• License IDs are publisher declarations, not legal clearance or proof of suitability. Review the model page and its actual terms before use. Model-card metadata.
• A model absent from a later bounded result list has left that query's returned sample. It may still exist. This is not a deletion detector or complete Hub monitor.

Scope and cost

The Actor reads public Hub API metadata only. It does not fetch weights, execute model code, call inference, collect private repositories or download model-card body text. It uses bounded public requests and reports source access failures explicitly; failures never silently become sample data. Hub API reference.

Free beta: no developer fee. Standard Apify compute, storage and platform usage apply. A run cap is a spending limit, not a price guarantee. For a small first run, use a $0.10 maximum total charge, 512 MB and a 300-second timeout. This Actor supports Limited permissions and does not require Standby mode or automatic restart.

For support, open an issue with a run URL, the query and the expected behavior. Remove credentials and private information before sharing logs.