Hugging Face Scraper: Models, Datasets & Spaces

Hugging Face Scraper exports Hugging Face Hub models, datasets, Spaces, daily papers and collections to CSV, Excel, JSON or XML. Search by keyword, author, task, library, language or license, filter by minimum downloads, likes or modified date, and get downloads, likes, trending score, gated status, benchmarks and inference providers per record. Optional README text and full model details. No login, no API key. Priced from $0.015 per model record.

It reads the Hub's public API, so numbers match huggingface.co. A search for "llama" sorted by trending returned meta-llama/Llama-3.1-8B-Instruct with 6,252,533 downloads and 8,271 likes in an earlier test run. It is built for AI engineers, analysts and compliance teams.

🤗 What does the Hugging Face Scraper do?

• Search models, datasets and Spaces by keyword, author or organization.
• Filter by task, library, language, license and tags, and by minimum downloads, likes or "modified since".
• Read daily papers (by term or date range) and collections with the items inside each.
• Add README / model card text (up to 100,000 characters) and, with one switch, license, parameter count, base model and file list.
• Return downloads, inference providers with price per million tokens, and benchmarks.
• Export as CSV, Excel, JSON or XML, or via the Apify API.

📊 What data can you extract from Hugging Face?

The dataset has 76 columns across five record types, and six views in the Output tab (Records overview, Models, Datasets, Spaces, Papers, Collections). A record only contains the fields that exist for it.

Record type	Main fields
Models	id, author, url, likes, downloads, downloadsAllTime, trendingScore, gated, pipelineTag, libraryName, languages, inferenceProviders, inferenceProviderDetails, benchmarks, arxivIds, createdAt, lastModified
Models with details	license, parameterCount, baseModel, fileCount, files
Datasets	taskCategories, sizeCategory, modalities, description, license
Spaces	sdk, hardware, appUrl, shortDescription, linked models and datasets
Papers	title, summary, authors, upvotes, publishedAt, githubRepo, githubStars, aiSummary, arxivUrl, pdfUrl
Collections	title, ownerFullname, itemsPreview, itemCount, itemTypeCounts, items (up to 200)
All types	recordType, searchTerm, scrapedAt, and readme when README text is on

Fields the Hub does not return are left out of the row. gated is No, auto or manual.

🧾 Sample output of a Hugging Face model search

Real row from cloud run M5gxukKdxCbcdr5Nn (earlier build, searchTerms: ["llama"], full details on). Newer builds add columns such as downloadsAllTime:

{
  "recordType": "model",
  "id": "meta-llama/Llama-3.1-8B-Instruct",
  "author": "meta-llama",
  "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct",
  "likes": 8271,
  "downloads": 6252533,
  "trendingScore": 69,
  "gated": "manual",
  "searchTerm": "llama",
  "pipelineTag": "text-generation",
  "libraryName": "transformers",
  "license": "llama3.1",
  "parameterCount": 8030261248,
  "files": [".gitattributes", "LICENSE", "README.md", "USE_POLICY.md"],
  "scrapedAt": "2026-10-07T20:04:50.773Z"
}

💵 How much does it cost to scrape Hugging Face?

Pay per event, only for records saved to the dataset. Failed requests, empty searches and error rows are never charged.

Event	What you get	Price (Free plan)
Model record	One model with downloads, likes, trending score, task, tags, gated status	$0.015 ($15 per 1,000)
Full model details	License, parameter count, base model and file list for one model	$0.010 ($10 per 1,000)
Dataset record	One dataset with downloads and likes	$0.01091 ($10.91 per 1,000)
Space record	One Space (demo app) with SDK and likes	$0.010 ($10 per 1,000)
Paper record	One paper with abstract, authors, upvotes, GitHub repository, AI summary	$0.008 ($8 per 1,000)
Collection record	One collection with owner, upvotes, item count and items	$0.008 ($8 per 1,000)
README text	Full README or model card text of one model, dataset or Space (only when found)	$0.012 ($12 per 1,000)

Gold plan: $0.01227 per model; other events are flat. Examples: 1,000 model records cost 1,000 x $0.015 = $15.00. 1,000 models with full details cost 1,000 x ($0.015 + $0.010) = $25.00. 1,000 models with README text cost 1,000 x ($0.015 + $0.012) = $27.00 when every README is found. 500 daily papers cost 500 x $0.008 = $4.00. "Max Items" and your maximum charge both stop the run cleanly.

🚀 How to scrape Hugging Face models in 3 steps

1. Type a search term (for example llama), an author or both.
2. Choose record types, add filters, set Max Items.
3. Click Start, then download the dataset as CSV, Excel, JSON or XML.

⚙️ Input

Field	Type	Default	What it does
searchTerms	list of text	empty	One Hub search per term (also papers and collections).
author	text	empty	Only records from this author or organization.
maxItems	number	10	Records to return in total. Free users: 10.
includeModels, includeDatasets, includeSpaces	boolean	true, false, false	Record types to scrape.
includePapers, papersStartDate, papersEndDate, papersSortBy	boolean, date, date, select	false	Papers by term (up to 100 per term) or daily feed; dates at most 200 days, end clamped to yesterday.
includeCollections, includeCollectionItems	boolean	false	Collections by search or owner; up to 200 items each.
pipelineTag, library, language, license, filterTags	select, text	empty	Hub tag filters (38 pipeline tags).
minDownloads, minLikes	number	0	Minimum 30-day downloads and likes.
modifiedFrom	text	empty	A date, or 7d, 4w, 3m, 1y, today, yesterday.
sortBy, sortDirection	select	trendingScore, desc	trendingScore, downloads, likes, lastModified or createdAt. asc only with createdAt.
includeModelDetails, includeReadme	boolean	false	License, parameters, base model, files; README text. Each billed as its own event.
hfToken	secret	empty	Optional Hugging Face token, checked before any request.

Example input, models with filters and README text:

{
  "searchTerms": ["embedding"],
  "library": "sentence-transformers",
  "language": "en",
  "minDownloads": 10000,
  "modifiedFrom": "3m",
  "includeReadme": true,
  "maxItems": 50
}

📦 Output

Each record is one dataset row. The Overview view shows the shared columns for all types; five other views show one record type each. A search that matches nothing finishes successfully with zero rows and zero charge.

⚖️ Hugging Face scraper vs alternatives

Read on 2026-10-10 from public Apify Store pages.

Capability	This actor	logiover	scrapers_lat
Surfaces	Models, datasets, Spaces, papers, collections	Same five	Models, datasets, Spaces
Filters	Library, language, license, tags, min downloads, min likes, modified date	Library, language, tags, min downloads, min likes, modified date	Task and sort only
README text	Models, datasets, Spaces, up to 100,000 chars	Yes, needs details on	Model card, paid plans only
Extra fields	Benchmarks, inference providers with prices, GGUF, 76 columns	About 24 fields	Benchmarks, providers, AI summary and tags (extra events)
Price per 1,000	$15 models (Free tier)	From $1.40	From $6.80

logiover covers the same five surfaces at a fraction of the price, with about 24 fields and no listed license filter, benchmarks or provider prices. scrapers_lat is closest on model data and bills model cards separately. seemuapps ($3.00 header) and hichemdev ($1.00 header) list models by query, task, library and sort only. We are the most expensive on model rows. Pick this actor when you need papers, collections, benchmarks, provider prices or the wider columns, and a cheaper one for plain lists.

🧩 Use cases

• Model selection with legal review: downloads and trending score for engineering, license and gated status for legal.

🔌 Run it via API, schedule it, connect it

Call it from Python or Node with the apify-client, or a plain HTTP request to the Apify API. Schedule a weekly trending sweep and connect the dataset to Zapier, Make, n8n, Google Sheets or Slack.

🤖 Hugging Face Scraper MCP server for AI agents

AI agents can call this actor through the Apify MCP server at https://mcp.apify.com?tools=RecordsData/huggingface-hub-scraper. The URL loads only this actor as a tool; add it to any MCP client that supports remote servers (Claude Desktop, Cursor, VS Code).

Example agent prompt:

Use the Hugging Face Scraper to list the 20 most downloaded text-generation models from mistralai, with full model details, and tell me which licenses they use.

Tool description: choose this actor when an agent needs Hugging Face Hub models, datasets, Spaces, daily papers or collections with downloads, likes, trending score, license, gated status, benchmarks or README text.

❓ Is it legal to scrape Hugging Face?

The actor reads only public information that the Hub's public API serves without login. It does not bypass access controls or read gated file contents. Respect each model's license. This is an independent tool, not affiliated with Hugging Face.

🙋 Frequently asked questions

How do I export Hugging Face models to CSV or Excel?

Enter a search term or an author in the Hugging Face Scraper, set Max Items, and click Start. The actor writes one row per model with downloads, likes, trending score and gated status. Then download CSV, Excel, JSON or XML.

How do I get the trending models on Hugging Face?

Leave "Sort by" on Trending in the Hugging Face Scraper, which is the default. Rows come back in the Hub's own trending order, and each row includes the trendingScore value. Add a task filter, such as text-generation, or an author to narrow it to one area.

How do I check the license of many Hugging Face models at once?

Switch on "Full model details" in the Hugging Face Scraper. Each model row then carries its license, parameter count, base model and file list, so you can review many licenses in one spreadsheet. Details are billed as their own event, only when data is returned.

What does the gated field mean in Hugging Face Scraper results?

In Hugging Face Scraper results, gated mirrors the Hub's gated-access setting. No means the repository can be downloaded directly, while auto or manual means you must request access on huggingface.co first. Gated models return only public metadata, never their gated files, so nothing private is read.

Does the Hugging Face Scraper also scrape datasets and Spaces?

Yes. Datasets and Spaces are separate switches in the Hugging Face Scraper, includeDatasets and includeSpaces, and both use the same search terms as models. Each surface is its own billing event, so you pay only for the types you switch on.

Can I list everything from one Hugging Face organization?

Yes. Set the author field, for example google, in the Hugging Face Scraper and leave the search terms empty. The actor returns records from that author or organization, sorted by your chosen order, up to Max Items. Free users get 10 records per run.

Why did the Hugging Face Scraper return 0 results?

The search matched no records on the Hugging Face Hub, so the Hugging Face Scraper ended successfully with zero rows and no charge. Try a shorter keyword, remove the task filter, or check the author spelling. If the Hub cannot be reached, the run fails with a clear message.

Do I pay for model details when the details call fails?

No. In the Hugging Face Scraper the "Full model details" event is charged only for models whose details call returned data. If the extra call fails, you keep and pay for the normal model record only. Failed requests are never charged.

How many records can one Hugging Face Scraper run return?

A paid-plan run of the Hugging Face Scraper can return up to 1,000,000 records, and stops cleanly at Max Items or at your maximum charge limit. Free Apify users get a 10-record preview per run, enough to check the output format.

What happens if I send an empty or invalid input to the Hugging Face Scraper?

An empty input {} is valid and returns the Hub's top 10 trending models. If the input cannot be run, for example when Models, Datasets and Spaces are all switched off, the run ends as SUCCEEDED with one free error row (errorType: invalid-input, with an error message and a hint). A bad sortBy value is rejected with a 400.

Does the Hugging Face Scraper need a Hugging Face token?

No. The Hugging Face Scraper uses public Hub data served without login, so you need only an Apify account, no Hugging Face account or API key. A token is optional through the hfToken input. Gated files are never read or downloaded, only public metadata is shown.

Are the filters applied by Hugging Face or by the actor?

Both. In the Hugging Face Scraper, library, language, license, task and filterTags are sent to the Hub as tag filters. minDownloads, minLikes and modifiedFrom are not supported by the Hub list, so the actor applies them to each page it reads.

Why does ascending sort only work with creation date?

The Hub accepts ascending order only for createdAt; any other sort answers with an error saying only descending sort is supported. The Hugging Face Scraper therefore stops before any request when you pick asc with another sort. Use createdAt with asc for the oldest repositories first.

Can I get the README or model card text, and what about gated repos?

Yes. Switch on includeReadme (billed as a README text event at $0.012 only when the text is found) and the Hugging Face Scraper adds a readme field (up to 100,000 characters) for models, datasets and Spaces. Gated repositories answer 401, so those rows come without the field.

Do I need a Hugging Face token (hfToken)?

Not for normal use. The optional hfToken input is a secret field, checked once before any list request, because the public endpoints silently ignore an invalid token. A rejected token ends the run with one free error row. The token is never logged. A valid token has not been tested in a cloud run.

Can it scrape Hugging Face papers and collections?

Yes. Turn on includePapers: with search terms you get up to 100 papers per term; without terms you get the daily feed, or one day at a time between papersStartDate and papersEndDate. Turn on includeCollections for collections by search or owner, and includeCollectionItems for up to 200 items each.

Does the actor return benchmarks and inference providers?

Yes, for models that have them. benchmarks holds the evaluation results published on the model page, inferenceProviders lists the Hub providers serving the model, and inferenceProviderDetails adds status, task, context length, price per million tokens and speed per provider. They add no extra call.

🆘 Support

Open an issue in the Issues tab or write to contact.punkrecordsdata@gmail.com.

Last updated: 2026-10-11