Hugging Face Scraper — Trending Models, Datasets & Spaces

Scrape trending Hugging Face models, datasets and Spaces in seconds: downloads, likes, trending score, task (pipeline tag), library, license, model size (parameters), base model, Space SDK and runtime status, plus the full model card / dataset card as Markdown. No Hugging Face account, token or browser needed.

Sort by trending, most downloaded, most liked, recently updated or newest, filter by author, task, library, minimum downloads/likes and release date, and turn on monitor mode to get only new trending models or likes & downloads growth since the last run. Schedule it daily and you have an AI trend feed.

Why this Hugging Face scraper

• All three Hub types: models, datasets and Spaces in one run, each with its own type-specific fields.
• Trending like on huggingface.co: the same trendingScore the Hub uses for its Trending pages.
• Rich model data: pipeline task, library (transformers, diffusers, gguf, mlx...), license, languages, base model, parameter count (12B, 744.4M) from safetensors/GGUF metadata, GGUF architecture and context length, downloads in the last 30 days and all time, gated status, commit sha, file count.
• Datasets: task categories, modalities, formats, size category, license, short description.
• Spaces: SDK (gradio, streamlit, docker, static), runtime stage (RUNNING, SLEEPING, RUNTIME_ERROR...), hardware, app URL, models and datasets used.
• Model cards for AI / RAG: optional README text with the YAML header removed, ready for LLMs and embeddings.
• Monitor mode: only new items, or only items whose likes/downloads grew, with growth numbers and percentages.
• Fast and polite: official public Hub API, 100 items per request, cursor pagination, respects the Hub's rate-limit headers and retries with backoff on HTTP 429.
• Pay per item: no monthly rental. In monitor mode you pay only for new or growing items.

Use cases

• AI market research: which tasks, model families, licenses and libraries are rising; who publishes the most downloaded open-weight models; how fast new releases get adopted.
• Model discovery: find the best trending text-generation, text-to-image, speech or embedding models, filtered by library (e.g. gguf or mlx for local inference) and minimum downloads.
• Investor / VC tracking: watch AI labs and startups (authors) for new releases and traction (likes and download growth) before it shows up in the news.
• AI newsletters & social posts: a daily or weekly list of newly trending models, datasets and Spaces with descriptions from their model cards.
• AI agents & RAG: feed model cards and metadata to an LLM agent that recommends models, writes summaries or answers "what is the best open model for X?".
• Competitive intelligence: track your own and competitors' models and Spaces (likes, downloads, Space status).
• Dataset sourcing: find trending datasets by task category with license and size information.

Input

Field	Description
resourceTypes	models, datasets and/or spaces (default ["models"]).
sort	trendingScore (default), downloads, likes, lastModified, createdAt. Always descending. Spaces have no downloads, so downloads sorts Spaces by likes.
maxItems	Max items per resource type (default 50, up to 10,000).
search	Keyword in the repo name, e.g. llama, whisper, flux.
authors	Users/organizations, e.g. google, meta-llama, Qwen.
pipelineTags	Tasks, e.g. text-generation, text-to-image. For datasets this is the task category. Ignored for Spaces.
libraries	Model libraries, e.g. transformers, diffusers, gguf, mlx. Models only.
minDownloads	Minimum downloads in the last 30 days (models and datasets).
minLikes	Minimum likes.
createdWithinDays	Only repos created in the last N days.
itemIds	Specific repos by id or URL: meta-llama/Llama-3.1-8B-Instruct, datasets/openai/gsm8k, spaces/enzostvs/deepsite, https://huggingface.co/....
includeCard	Add the README / model card text (Markdown, YAML header removed).
cardMaxChars	Max card length (default 20,000 characters).
monitorMode	off (default), new, growth or newAndGrowth.
monitorId	Separate saved state per watchlist (default default).
baselineOnFirstRun	First monitored run saves items silently (no output, no cost). Default on.
minLikesGrowth	Growth mode: minimum likes gained since the previous run (default 1).
minDownloadsGrowth	Growth mode: minimum downloads gained (all-time downloads when available). 0 = ignore downloads.

Several authors, tasks and libraries are combined (every combination is one Hub query, the results are merged, de-duplicated and re-sorted, at most 100 combinations).

Example input: today's trending models, datasets and Spaces

{
  "resourceTypes": ["models", "datasets", "spaces"],
  "sort": "trendingScore",
  "maxItems": 20,
  "includeCard": true
}

Example input: new local-inference LLMs from the last 7 days

{
  "resourceTypes": ["models"],
  "sort": "trendingScore",
  "pipelineTags": ["text-generation"],
  "libraries": ["gguf", "mlx"],
  "createdWithinDays": 7,
  "minDownloads": 1000,
  "maxItems": 100
}

Example input: daily monitor of AI labs (new releases + traction)

{
  "resourceTypes": ["models", "spaces"],
  "sort": "createdAt",
  "authors": ["google", "meta-llama", "Qwen", "mistralai", "deepseek-ai"],
  "maxItems": 200,
  "monitorMode": "newAndGrowth",
  "monitorId": "ai-labs",
  "minLikesGrowth": 25
}

Output

One row per model, dataset or Space. Real example (model, from a test run on 2026-10-14; tags and card shortened for readability):

{
  "type": "model",
  "id": "jialinyyzz/humanizer",
  "name": "humanizer",
  "author": "jialinyyzz",
  "url": "https://huggingface.co/jialinyyzz/humanizer",
  "title": null,
  "pipelineTag": "text-generation",
  "library": "gguf",
  "tags": ["gguf", "safetensors", "gemma4_unified", "image-text-to-text", "humanizer", "transformers", "text-rewriting", "llama.cpp", "gemma4", "text-generation", "en", "zh", "base_model:google/gemma-4-12B", "base_model:finetune:google/gemma-4-12B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational"],
  "license": "apache-2.0",
  "languages": ["en", "zh"],
  "downloads": 65616,
  "downloadsAllTime": 67722,
  "likes": 1571,
  "trendingScore": 1072,
  "createdAt": "2026-09-11T05:00:16.000Z",
  "lastModified": "2026-10-08T20:13:31.000Z",
  "gated": false,
  "private": false,
  "disabled": false,
  "sha": "7232532cb8466bd393544e97826e12e6b5e471ac",
  "baseModels": ["google/gemma-4-12B"],
  "datasets": [],
  "parameters": 11959730224,
  "parametersHuman": "12B",
  "parameterDtypes": ["BF16"],
  "ggufArchitecture": "gemma4",
  "contextLength": 262144,
  "filesCount": 38,
  "card": "<img src=\"assets/banner-en.png\" alt=\"humanizer: rewrites AI drafts so they read like a person wrote them\" width=\"100%\">\n\n...\n\n# humanizer\n\n**A 12B model that rewrites AI-written drafts (emails, essays, reports, forum posts; English and Chinese) so they read like a person wrote them.** It is trained to keep every number, unit, date",
  "cardError": null,
  "cardTruncated": true,
  "scrapedAt": "2026-10-14T09:53:31.950Z"
}

Space (same run):

{
  "type": "space",
  "id": "Pepe104/MiniMax-H3-Turbo-Lora-UNCENSORED",
  "author": "Pepe104",
  "url": "https://huggingface.co/spaces/Pepe104/MiniMax-H3-Turbo-Lora-UNCENSORED",
  "title": "MiniMax H3 Turbo LoRA",
  "likes": 1526,
  "trendingScore": 239,
  "createdAt": "2026-08-14T23:23:19.000Z",
  "emoji": "🎬",
  "shortDescription": "Video generation with a synchronized soundtrack",
  "sdk": "gradio",
  "sdkVersion": "6.20.0",
  "runtimeStage": "RUNNING",
  "hardware": "zero-a10g",
  "appUrl": "https://pepe104-minimax-h3-turbo-lora-uncensored.hf.space",
  "models": ["hfmlsoc/ncii-light-guard-v01", "larryvrh/MiniMax-H3-Turbo-Lora", "lightx2v/Minimax-h3-Turbo", "MiniMaxAI/MiniMax-H3", "multimodalart/minimax-h3-aoti"]
}

Datasets add title (pretty name), taskCategories, modalities, formats, sizeCategory and description. Fields that do not apply to a type are null (e.g. downloads for Spaces).

In monitor mode each row also has:

{
  "changeType": "growth",
  "likesGrowth": 7,
  "likesGrowthPercent": 0.3,
  "downloadsGrowth": 5000,
  "downloadsGrowthBasis": "allTime",
  "previousLikes": 2150,
  "previousDownloads": 124930,
  "previousDownloadsAllTime": 119930,
  "previousSeenAt": "2026-10-13T08:00:00.000Z",
  "firstSeenAt": "2026-10-01T08:00:00.000Z"
}

changeType is new (not seen in previous runs) or growth. A run report (queries, items scanned and matched, warnings, errors, not-found repos) is saved to the SUMMARY record of the run's key-value store.

How monitoring works

• State is saved in the named key-value store huggingface-monitor under state-<monitorId> (gzipped JSON).
• With baselineOnFirstRun (default), the first run of a list (resource type + sort + filters) saves its items silently. Later runs report items that are new to the list (e.g. newly trending) or that gained likes/downloads.
• Growth is measured since the previous run that saw the item. Downloads growth uses all-time downloads when the Hub provides them (models and datasets), because the 30-day figure can go down.
• If the run's cost limit is reached, unreported items stay pending and are reported in the next run.
• Items not seen for 180 days are dropped from the state.

Pricing

Pay per item: you pay a small fee for each model, dataset or Space in the dataset (event item). Model cards are included at no extra charge. In monitor mode you pay only for new or growing items, so a daily trending watch usually costs cents. Set a maximum cost per run in the run options and the Actor stops cleanly when it is reached.

FAQ

Do I need a Hugging Face account or API token? No. The Actor uses the public Hugging Face Hub API and only reads public repos.

What is the trending score? The Hub's own popularity score for recent activity (likes and usage over the last days). It is what huggingface.co uses for its Trending sort. trendingScore is the value at scraping time.

Why are downloads empty for Spaces? The Hub does not count downloads for Spaces. Use likes, trendingScore and runtimeStage instead. minDownloads is ignored for Spaces and sort: downloads sorts them by likes.

Why is card empty for some models? Gated models (e.g. Llama, Gemma with license acceptance) do not publish their README to anonymous users: cardError is gated or private. Repos without README get no README, repos with only a YAML header get empty.

How are minDownloads, minLikes and createdWithinDays applied? The Hub API cannot filter on them, so the Actor reads the sorted list and skips items that do not match. With a matching sort (downloads, likes or createdAt) it stops at the first item below the threshold. With another sort it scans up to 10,000 items per query, so very strict filters may return fewer items (see SUMMARY).

Can I scrape one specific model? Yes, put its id or URL in itemIds (for example Qwen/Qwen-Image-2.1 or https://huggingface.co/datasets/openai/gsm8k). You can leave resourceTypes empty to scrape only those repos.

How fresh is the data? Live: every run calls the Hub API. Download counts are the Hub's own (updated by the Hub about daily).

Limitations

• Only public repos. Private repos are not visible; gated repos are listed but their cards are not readable.
• The Hub API combines filters with AND per query, so several authors/tasks/libraries are queried one by one (max 100 combinations per run).
• pipelineTags and libraries do not apply to Spaces; libraries does not apply to datasets.
• The Hub limits anonymous API use (about 500 requests per 5 minutes). The Actor pauses when the limit is near, so very large runs (tens of thousands of items) take longer.

Legal

The Actor reads only public metadata and README files that Hugging Face shows to every visitor through its public API. It does not log in, bypass gating or collect personal data beyond public usernames of repo owners. Model cards, datasets and code remain the property of their authors and are subject to their licenses (see license). Make sure your use of the data complies with the Hugging Face Terms of Service and applicable laws.

Related tools

• Substack to Markdown for AI & RAG
• Career Sites Jobs Feed (11 ATS)
• Shopify Store Products & Price Change Monitor