Hugging Face Model Intelligence

What does Hugging Face Model Intelligence do?

Hugging Face Model Intelligence collects public model data from the official Hugging Face Hub API and returns it as a clean dataset you can export or schedule. Filter by library, task, license, language and search text, sort by downloads, likes, last update or trending score, and get one row per model with:

• model id and public URL
• library, task (pipeline tag), declared license and card languages
• downloads, likes and the upstream trendingScore
• last modified date
• an optional short plain-text model card excerpt when the API explicitly reports an ungated repository and a supported standard repository license declaration, with common contact patterns removed

No Hugging Face login, no browser, no paid proxy, no inference service, and no model weights are downloaded.

Why use it?

• Spot rising models early: sort by trendingScore and run daily to see what the community is picking up this week.
• Shortlist by license: list apache-2.0 or mit text-generation models in a given language before you commit to a stack.
• Market and investor research: track adoption (downloads, likes) of a library or task over time by scheduling snapshots.
• Competitive and ecosystem monitoring: watch new and recently updated models for a task such as text-to-image or automatic-speech-recognition.
• Feed your own tools: push the dataset to Google Sheets, a database, Slack alerts or an LLM pipeline through Apify integrations and webhooks.

Downloads and likes are upstream popularity signals, not verified users, revenue or quality, and nothing here is investment advice.

How to collect Hugging Face models

1. Open the Input tab.
2. Set optional filters, for example library transformers, task text-generation, license apache-2.0, language en. The Hub API combines them; leave a filter empty to broaden the search.
3. Pick a Sort by value and bound the work with Max models and Max API pages.
4. Click Start, then download the dataset as JSON, CSV, Excel or HTML. The SUMMARY record in the key-value store shows counts, errors and why the run stopped.

Input

Field	Default	What it does
search	empty	Text matched against model names.
library	empty	Library tag such as transformers, diffusers, gguf.
task	empty	Task tag such as text-generation, text-to-image.
license	empty	License id without the license: prefix, such as apache-2.0.
language	empty	Language tag such as en or tr.
maxModels	20	1 to 10,000 model rows.
pageSize	20	Models per API page, 1 to 100.
maxPages	10	Safety limit on API pages, 1 to 1,000.
sort	downloads	downloads, likes, lastModified or trendingScore, always descending.
includeCardSummary	true	Fetch an eligible public README only (see rights policy below). Turn off for metadata-only runs.
summaryMaxChars	600	Excerpt length, 100 to 2000.
requestDelayMillis	500	Pause before every request, 100 to 10,000.

Pagination follows the API's cursor Link headers and duplicate model ids are removed. Example input:

{
  "library": "transformers",
  "task": "text-generation",
  "license": "apache-2.0",
  "language": "en",
  "sort": "trendingScore",
  "maxModels": 5
}

Output

Each dataset row is one model. Real local result on 2026-10-22 for input {"search":"all-MiniLM-L6-v2","maxModels":2} (counts change over time, excerpt shortened):

{
  "modelId": "sentence-transformers/all-MiniLM-L6-v2",
  "modelUrl": "https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2",
  "library": "sentence-transformers",
  "task": "sentence-similarity",
  "license": "apache-2.0",
  "languages": ["en"],
  "downloads": 198627085,
  "likes": 5996,
  "trendingScore": 22,
  "lastModified": "2026-06-01T06:29:13.000Z",
  "cardSummary": "all-MiniLM-L6-v2 This is a sentence-transformers model: It maps sentences & paragraphs to a 384 dimensional dense vector space ...",
  "cardStatus": "available",
  "fetchedAt": "2026-10-22T18:14:52.442Z"
}

Field	Type	Meaning
modelId, modelUrl	string	Repository id and public page
library, task, license	string or null	Upstream declarations
languages	string array	Languages declared on the model card
downloads, likes	number or null	Current Hub counts
trendingScore	number or null	Upstream score; never invented when absent
lastModified	string or null	Hub timestamp
cardSummary	string or null	Truncated README excerpt, not an AI summary
cardStatus	string	available, unavailable (404), error, disabled, restricted or rights-unclear
fetchedAt	string	UTC collection time

How much does it cost?

The Actor uses pay-per-event pricing: $0.0015 per model row ($1.50 per 1,000 models), with no start fee. Rows for valid models whose card is unavailable still count as one result; error details go only to the SUMMARY record and are not billed. For comparison, other Hugging Face scrapers on the Store listed about $1 to $3 per 1,000 results when checked on 2026-10-22. The Pricing tab on the Store page is authoritative for the current price. Use the maximum cost per run setting to cap spend; the Actor stops cleanly when the limit is reached. Apify platform usage is included in the event price.

FAQ

Which API does it use? The public Hub API (/api/models) for the list and the public raw/main/README.md file of each model for the card excerpt. Nothing private is requested.

Does it cover datasets or Spaces? Not yet. This version returns models only.

What is trendingScore? The value Hugging Face reports for how fast a model is gaining attention. Sort by it to see what is rising. It is null when the API does not return one.

Why is cardSummary empty for some models? A missing README (404) is unavailable. Gated or unknown gate status is restricted; an absent, custom or unsupported license declaration is rights-unclear. These skipped cards are not requested, but their public API metadata can still be returned. HTTP 401/403/429 and detected access challenges stop the run before the affected row or any further request. Models with an empty eligible README return null.

Does it collect personal data? No. Model cards are public text, but common e-mail and phone patterns and URL targets are removed from excerpts. The Actor is not designed to collect contact information.

What happens on errors or rate limits? HTTP 401/403/429 and detected access challenges (including challenge pages returned as HTTP 200 or 503) fail the run immediately without retry, fallback, further public requests or emission of the affected row. Existing rows and deduplication state are retained; STATE records access-denied and SUMMARY records the failure. Ordinary card 404 and other non-denial failures remain partial results; list failures fail the run. Transport errors and non-challenge 5xx responses are retried once. Each request has a 20-second timeout.

Can I schedule it? Yes. Use Apify schedules to snapshot downloads and trending scores daily or weekly. Each run starts a fresh collection; the same run restores its STATE checkpoint after a restart and skips previously emitted models. STATE is saved at page boundaries, charge limits, access denials and on abort.

Is it legal? It reads public Hub endpoints. Public registry metadata is not a download or license grant for model weights, and metadata eligibility does not require an open weight license. Private models are skipped; no private/gated card or weights endpoint is requested.

For optional card text, the conservative policy requires gated: false and a declared repository license of mit, apache-2.0, bsd-2-clause, bsd-3-clause, cc0-1.0 or cc-by-4.0, without custom license_name/license_link overrides. Other cards are not downloaded or excerpted. This is a declaration-based filter, not verification that every README is licensed identically to the model. Keep the returned license, source modelUrl and card status with any excerpt; verify the source license, notices, attribution and modification requirements before further redistribution. Set includeCardSummary: false for metadata only.

The Hugging Face terms retain accompanying license conditions and references; the model-card documentation describes declared and custom license metadata. Accessibility alone is not a blanket reuse license. Declared licenses and languages can be incomplete and are not legal advice.

Can I call it from code? Yes, through the API tab or the Apify client libraries for JavaScript and Python.

Local development and support

Node.js 22 or later: npm ci --ignore-scripts, npm test (offline behavioral tests), apify validate-schema, and apify run --input-file input.json. Dependencies are pinned in the lockfile; an indirect SDK FTP dependency is overridden to basic-ftp 6.2.3 for GHSA-c475-qrg2-pj4r.

Report reproducible problems in the Store Issues tab with the input you used. This Actor is independent and is not affiliated with or endorsed by Hugging Face.