Hugging Face Scraper

🤗 Hugging Face Scraper

Export models, datasets and spaces from the Hugging Face Hub in seconds: downloads, likes, trending score, task, library, license, parameter count, base models, languages, tags, files and the full model card. Search by keyword, author, task, library or tag, or paste any Hub URL. No login, no API token, $0.50 per 1,000 repos.

✨ Why use Hugging Face Scraper?

• ⚡ Fast: reads the Hub's own JSON API, up to 1,000 repos per request. A 10,000-model run finished in about 16 seconds in our tests.
• 🧩 All three repo types: models, datasets and spaces, each with its own fields (parameters and base models, dataset formats and size, space SDK, hardware and live status).
• 🔎 Every way in: keyword search, author or organization, task, library, language, license, tags, collections, or a listing URL copied straight from the website with its filters.
• 📈 Trending and top lists: sort by trending, downloads, likes, recently created or recently updated.
• 📝 Deep details on demand: architecture, spaces using a model, fine-tune count, file list and the README card text.
• 🔓 No account needed: only public Hub data, no token, no cookies.

🎯 Use cases

• 📊 Market and trend research: track which open models and datasets are rising each week, by task or by organization.
• 🏢 Competitive intelligence: monitor everything a lab or company (meta-llama, mistralai, Qwen, google) publishes, with downloads and likes over time.
• 🧪 Model selection: shortlist models for a task by license, size (parameters), library and popularity.
• 📚 Dataset discovery: find datasets by task, language, modality, format or size for training and evaluation.
• 🗂️ Catalogs and dashboards: feed an internal model registry, a newsletter or a BI dashboard on a schedule.
• 🎓 Academic research: study the open-source AI ecosystem: licenses, base-model lineage, arXiv links and adoption.

🚀 How to use

1. Pick What to scrape: models, datasets or spaces.
2. Enter search queries (for example llama, whisper), and/or paste Hugging Face URLs: repo pages, author pages, collections or filtered listing pages.
3. Optionally narrow it down with Filters (author, task, library, language, license, tags, min downloads or likes, dates) and choose a sort order.
4. Set Max results per search and run.
5. Download the results as JSON, CSV, Excel or HTML, or read them with the API.

📥 Input

{
  "type": "models",
  "searchQueries": ["llama"],
  "startUrls": [
    "https://huggingface.co/openai/gpt-oss-20b",
    "https://huggingface.co/models?pipeline_tag=text-to-image&library=diffusers&sort=downloads"
  ],
  "pipelineTag": "text-generation",
  "license": "apache-2.0",
  "sort": "downloads",
  "maxResults": 20,
  "includeDetails": false,
  "includeReadme": false
}

📤 Output

Each repo is one row. Common fields come first, then the fields for that repo type.

Example (trimmed):

{
  "type": "model",
  "id": "meta-llama/Llama-3.1-8B-Instruct",
  "name": "Llama-3.1-8B-Instruct",
  "author": "meta-llama",
  "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct",
  "likes": 8415,
  "trendingScore": 62,
  "downloads": 6034692,
  "downloadsAllTime": 189862190,
  "pipelineTag": "text-generation",
  "libraryName": "transformers",
  "parameters": 8030261248,
  "parametersByDtype": {"BF16": 8030261248},
  "baseModels": ["meta-llama/Llama-3.1-8B"],
  "baseModelRelation": "finetune",
  "languages": ["en", "de", "fr", "it", "pt", "hi", "es", "th"],
  "arxivIds": ["2204.05149"],
  "license": "llama3.1",
  "tags": ["transformers", "safetensors", "llama", "text-generation", "conversational", "license:llama3.1"],
  "gated": "manual",
  "createdAt": "2024-07-18T08:56:00.000Z",
  "lastModified": "2024-09-25T17:00:57.000Z",
  "source": "llama"
}

A space row looks like this:

{
  "type": "space",
  "id": "enzostvs/deepsite",
  "title": "DeepSite v4",
  "emoji": "🐳",
  "shortDescription": "Generate any application by Vibe Coding it",
  "sdk": "docker",
  "runtimeStage": "RUNNING",
  "hardware": "cpu-xl",
  "appUrl": "https://enzostvs-deepsite.hf.space",
  "likes": 16610,
  "models": ["deepseek-ai/DeepSeek-V3.2", "Qwen/Qwen3-Coder-30B-A3B-Instruct"]
}

💰 Pricing

You pay $0.50 per 1,000 repos saved ($0.0005 per repo), plus a small amount of platform usage (compute and datacenter proxy traffic), which for this scraper is usually a fraction of a cent per run.

• 1,000 models ≈ $0.50
• 10,000 models ≈ $5.00

The detail, file list and README options cost nothing extra per repo; they only make the run take longer. You are never charged for repos that were not saved, and the run stops cleanly when it reaches your maximum charge.

❓ FAQ

Is it legal to scrape Hugging Face? The scraper reads only public data that the Hub serves to anyone without logging in, through its public API. Check the license of each model or dataset before you use the files themselves.

Do I need a Hugging Face account, token or proxy? No. It works without any account. Datacenter proxies are the default; the scraper switches to a fresh IP when the Hub's anonymous rate limit is reached.

How many results can I get? Up to 200,000 per search. The Hub's own search matches repo names, so broad words return many results; add filters to narrow them.

Why are some fields empty? Fields only appear when the repo has them: spaces have no downloads, many models have no safetensors parameter count, and readme, files and the detail fields are only filled when you turn those options on.

Can I get private or gated files? No. Private repos are not visible without an account. Gated repos (marked with gated) are listed with all public metadata, but their files stay behind the gate.

Can I schedule it or use it from code? Yes. Schedule runs in the console, or start runs and fetch results through the API and integrations.

⚠️ Limits

• Search matches repo names (like the Hub's search box), not README text.
• Download counts are the Hub's rolling last-30-days figure plus an all-time total; there is no daily history.
• Papers, posts, discussions and user profiles are not scraped. Collections are supported for the models, datasets and spaces they contain.
• README text is capped at 50,000 characters per repo.