Hugging Face Hub Search: Models, Datasets, Spaces

Hugging Face Hub search that returns one flat row per model, dataset, Space or daily paper: downloads, likes, task, license, language, tags and dates, ready for a spreadsheet, a database or an AI agent. Pick a resource type, add search text or filters, set a maximum and run. Results cost $3.00 per 1,000 rows, and a run of 20 rows costs about six cents. I built this Actor as an independent developer. It is not affiliated with or endorsed by Hugging Face.

Sample output

One row from a Hugging Face API model search (task text-generation, sorted by downloads). Values were read from the Hub on 2026-10-04 and change over time. Fields with no value (null) are left out here for space.

{
  "rank": 1,
  "recordType": "model",
  "id": "Qwen/Qwen3-0.6B",
  "owner": "Qwen",
  "name": "Qwen3-0.6B",
  "url": "https://huggingface.co/Qwen/Qwen3-0.6B",
  "task": "text-generation",
  "library": "transformers",
  "license": "apache-2.0",
  "downloads": 29665992,
  "downloadsAllTime": 234340865,
  "likes": 1727,
  "trendingScore": 31,
  "parameters": 751632384,
  "gated": false,
  "createdAt": "2025-04-27T03:40:08.000Z",
  "lastModified": "2025-07-26T03:46:27.000Z",
  "tags": ["transformers", "safetensors", "qwen3", "text-generation", "license:apache-2.0"]
}

How to search the Hugging Face Hub with this Actor

1. Open the Actor and choose a Resource type: Models, Datasets, Spaces or Daily papers.
2. Type Search text to match repository names, for example llama or sentiment. Leave it empty to list the top results for your sort order. For papers, type a topic such as diffusion.
3. Set Maximum results. The prefill is 20, so the first run is cheap. One run returns up to 10,000 rows.
4. Open Filters to narrow by owner (for example meta-llama), task, license, language, library or Space SDK.
5. Open Advanced options to change the sort order, add raw Hub tags or pick a paper date.
6. Click Start. Rows appear in the dataset as they are saved. Use the Overview table, or export to JSON, CSV or Excel.

You can also call the Actor from code through the Apify API. The example below lists the 100 most liked Gradio Spaces.

curl -X POST "https://api.apify.com/v2/acts/conserving_celerytop~huggingface-hub-search-api/run-sync-get-dataset-items" \
  -H "Authorization: Bearer <YOUR_APIFY_TOKEN>" \
  -H "Content-Type: application/json" \
  -d '{"resource": "spaces", "spaceSdk": "gradio", "sortBy": "likes", "maxItems": 100}'

What data you get from the Hugging Face API

Field	Applies to	Meaning
rank, recordType, id, url	all	Position in the result, type, repository id such as owner/name, and the Hub link
owner, name	models, datasets, Spaces	The namespace and repository name from the id
task, library, parameters	models	Pipeline tag, library name and total parameter count when the repository lists it
task, description	datasets	First task category (or the one you filtered by) and a short cleaned description
license, languages, tags	models, datasets, Spaces	License id, language codes and the full tag list
downloads, downloadsAllTime	models, datasets	Downloads in the last 30 days and over all time
likes, trendingScore	models, datasets, Spaces	Likes and the Hub trending score
gated	models, datasets	false, auto or manual access setting
sdk, runtimeStage, hardware	Spaces	SDK, current stage and hardware name
title, description, upvotes, numComments, organization, githubRepo, githubStars, projectPage, publishedAt, arxivUrl, aiSummary, keywords	papers	Paper title, abstract, community counts, linked code and project page
createdAt, lastModified, fetchedAt	all	Hub dates and the time the row was read

Fields that do not apply to a resource type are null or an empty list, so every row has the same columns. When a repository lists several licenses or task categories, the one you filtered by is shown, and all of them stay in tags.

Hugging Face Hub search pricing

Pricing is pay per event. One event is one saved result row, plus the standard Actor start event.

Plan	Price per result	Price per 1,000 results
Free	$0.0030	$3.00
Bronze	$0.0027	$2.70
Silver	$0.0024	$2.40
Gold	$0.0021	$2.10

Worked example: a search for the 500 most downloaded text-generation models returns 500 rows. On the Free plan that is 500 x $0.003 = $1.50, plus $0.00005 for the Actor start, so about $1.50 in total. The same run on the Gold plan costs $1.05.

How charging works:

• A row is charged only after it is saved to the dataset.
• A valid search with no matches saves one message row and charges one result.
• A search the Hub rejects returns one error row with the Hub message and charges one result.
• Input the Actor rejects before any request (for example a Maximum results of 0) ends the run with a plain message and charges nothing.
• When Hugging Face is unavailable or its rate limit is used up and no rows were saved, the run saves one error row and charges one result. Rows saved before such a stop are charged as usual.
• The run stops when the maximum charge you set for the run is reached.

Input example

{
  "resource": "datasets",
  "searchQuery": "sentiment",
  "task": "text-classification",
  "license": "mit",
  "language": "en",
  "sortBy": "downloads",
  "maxItems": 200
}

Field	Type	Notes
resource	select	models, datasets, spaces or papers
searchQuery	text	Matches repository names. For papers it runs the Hub paper search
maxItems	integer	1 to 10,000. A paper search returns up to 120
owner, task, license, language, library, spaceSdk	text or select	Filters that combine with AND
sortBy	select	downloads, likes, trendingScore, createdAt, lastModified. The default is downloads for models and datasets, likes for Spaces, the Hub's newest-first order for papers
tags	list	Raw Hub tags such as size_categories:1M<n<10M
paperDate	text	YYYY-MM-DD, for the daily papers feed
onlyNew	checkbox	Return only items this monitor has not delivered before. The first run returns everything and becomes the baseline. See Run it every day
monitorName	text	Names the memory of the monitor. Use one name per saved task. Letters, numbers, - and _, up to 40 characters

Output example for a daily paper

{
  "rank": 1,
  "recordType": "paper",
  "id": "2609.22753",
  "url": "https://huggingface.co/papers/2609.22753",
  "title": "Replacing Large Language Models with Jev Decision Models for Low-Latency Edge Service Orchestration",
  "upvotes": 5,
  "numComments": 2,
  "githubRepo": "https://github.com/OniReimu/Edge-Computing-JEV",
  "publishedAt": "2026-09-26T00:00:00.000Z",
  "arxivUrl": "https://arxiv.org/abs/2609.22753"
}

Paper rows leave out author names, submitter profiles and avatars.

Use cases for a Hugging Face model and dataset list

• Build a catalog of open models by task, license and library, for example every apache-2.0 text-generation model with more than 100,000 downloads.
• Track trending models, datasets and Spaces each week with a scheduled run sorted by trending score.
• Compare model families by downloads, likes and parameter count.
• Collect dataset candidates for a task and language, with license and size tags.
• Feed daily papers with their linked code repositories into a reading list or an AI agent.

Run it every day

Switch on Only new since the last run and the Actor returns only models, datasets, Spaces or papers that it has not delivered before.

1. Fill the input, for example your search text and filters. Set Sort by to Newest created, so new repositories come first.
2. Switch on Only new since the last run and enter a Monitor name, for example new-llama-models. Use one name per saved task.
3. Click Save as task and run it once. The first run returns up to Maximum results items and becomes the baseline.
4. In the Apify Console, open Schedules, click Create new, pick your Task, and enter a cron expression for once a day, for example 0 7 * * *.

Each later run returns only items that are not in the memory. Items skipped as already seen are not charged and do not count toward Maximum results. Items that were cut by Maximum results or by your maximum charge are not remembered, so a later run delivers them. If a search matches more items than Maximum results, the next run continues with the older items you have not received yet. Set Maximum results above the number of new items you expect per day.

The Actor stops reading after 500 items in a row that it delivered before. A run with nothing new reads one page of 500 items and ends with the message "No new items since the last run".

The memory lives in your own key-value store, named hf-monitor-<monitor name>. It holds ids only, such as model|meta-llama/Llama-3.1-8B. If you change the resource type, search text or any filter, the Actor starts a separate memory for that search, and its first run returns everything again. Maximum results is not part of the memory key.

Worked cost example on the Free plan, with the prices in the pricing table above ($0.003 per row, $0.00005 for the start). Say 40 new models match your search each day. A day costs 40 x $0.003 = $0.12 plus $0.00005, so $0.12005. Thirty days cost 30 x $0.12005 = $3.6015. A day with no new model costs only the start fee of $0.00005. The same search run daily without Only new, with Maximum results 1,000, returns up to 1,000 rows every day: 1,000 x $0.003 = $3.00 plus $0.00005 per day, and 30 days cost 30 x $3.00005 = $90.0015. The 40 new models per day is an example, not a measured value.

Related Actors

• Crossref Works Search finds papers, DOIs and citation counts from the Crossref index.
• AI Jobs Search lists LLM and machine learning jobs from company career pages.

FAQ

Is it legal to list Hugging Face models and datasets with this Actor?

The Actor reads the public Hub API without a login. It returns metadata only: repository names, counts, tags, license ids and dates. It does not download model files or dataset contents, it does not read private repositories, and it does not collect user profiles. The Hugging Face Terms of Service say that content carrying a license "is intended to remain under the terms of such license when further accessed, distributed, or used". The license field shows what each repository declares, so check it before you use a model or dataset. This is not legal advice.

Which Hugging Face API endpoints does it use?

The public list endpoints for models, datasets and Spaces, plus the daily papers and paper search endpoints. It sends one request per page of up to 1,000 rows and follows the next-page link the Hub returns.

What are the limits?

• Up to 10,000 rows per run. Start another run with a narrower filter or a different sort to go further.
• A paper search returns up to 120 rows. The daily papers feed has no such cap.
• Search text matches repository names, not README text.
• The Hub applies an anonymous quota of 500 API calls per 5 minutes per IP address. A 10,000-row run uses 10 calls.
• downloads counts the last 30 days. downloadsAllTime counts every download.

Why did my run stop early or return fewer rows?

Open the STATS record in the key-value store. The status field says why: succeeded, no_results, invalid_input, rejected_by_source, rate_limited, source_unavailable or charge_limit_reached. The status message on the run page repeats it in a sentence. If the Hub rate limit was used up for longer than the Actor waits (100 seconds), the rows saved so far are kept and a new run a few minutes later starts with a fresh quota.

Why do I get a row with recordType set to no_results?

The search matched nothing. Loosen the filters or shorten the search text. Filters combine with AND, so each extra filter can only reduce the list.

A filter was ignored. Why?

Some filters apply to one resource type only, for example Space SDK applies to Spaces and Library applies to models. The STATS record lists the ignored filters in ignoredFilters.

Can I schedule it?

Yes. Use an Apify schedule with the same input, then compare runs by id to see which repositories are new.

If this saved you time, a rating helps others find it.