Hugging Face Model & Dataset Trend Tracker

Track trending Hugging Face models and datasets — downloads, likes, and velocity — straight from the official public Hugging Face Hub API. No browser, no scraping tricks, no API key required. Filter by task (text-generation, text-to-image, automatic-speech-recognition…), library (transformers, diffusers, gguf…), or tag, and switch on monitor mode to get alerted the moment a new LLM, model, or dataset starts trending.

This is AI-model trend intelligence, not a static catalog dump: instead of "list every model," it answers "what's gaining adoption on Hugging Face right now." Built for ML engineers, DevRel and LLM-ops teams, AI researchers, and data teams who need to see what's rising across the open-model ecosystem. (Pairs with our Civitai Model & Prompt Trend Tracker for the creative/AI-art side.)

What you get

Both result types land in one dataset, each row tagged with resultType ("model" or "dataset") so you can split them downstream.

Trending model fields

Field	Description
modelId, author, modelName	Full repo id (owner/name), owner, and short name
pipelineTag	Canonical task, e.g. text-generation, text-to-image (may be null if untagged)
libraryName	Library/framework, e.g. transformers, diffusers, gguf
tags	All Hub tags on the repo
downloads, likes	Adoption metrics
trendingScore	Hugging Face's own trending signal (null if not exposed for the chosen sort)
lastModified, createdAt	ISO 8601 timestamps
isPrivate, gated	Visibility and gating (false / "manual" / "auto")
cardData	Model-card front-matter: license, language, datasets, baseModel
siblings	Count of files in the repo (proxy for model size/complexity)
config	config.json architecture metadata — only when includeConfig is on
modelUrl	Direct Hugging Face link
trendRank	Position in the sorted results (1 = top)
isNew	true if new to trending since the last monitor run (null in snapshot mode)
scrapedAt	ISO 8601 timestamp of this run

Trending dataset fields

Field	Description
datasetId, author, datasetName	Full id (owner/name), owner, and short name
tags	All Hub tags
downloads, likes	Adoption metrics
trendingScore	Hugging Face trending signal (null if not exposed)
lastModified, createdAt	ISO 8601 timestamps
isPrivate, gated	Visibility and gating
cardData	Dataset-card front-matter: license, language, size, taskCategories
datasetUrl	Direct Hugging Face link
trendRank, isNew, scrapedAt	Trend + monitor metadata

Two data types — which to use

• Trending models (dataType: "models", the default) — what models and LoRAs/checkpoints are gaining adoption. Track model releases, discover emerging architectures, and watch a task or library heat up.
• Trending datasets (dataType: "datasets") — which training/eval datasets are rising. Discover new corpora and benchmarks before they're everywhere.
• Both (dataType: "both") — run both in one pass; rows are tagged with resultType.

Monitor mode — track new trends over time

Set mode: "monitor" and attach an Apify Schedule (daily or weekly):

• First run stores the current trending IDs and returns everything.
• Later runs return only entries new to the trending list since the previous run, each flagged isNew: true.
• If nothing new is trending, the run finishes cleanly with an empty dataset (not an error).

State is keyed per data type and sort (state-models-trending, state-datasets-downloads, …) and persisted in a named Key-Value Store, so it survives across scheduled runs. This is the "alert me when a new model starts trending" workflow — feed it into Slack, email, or a webhook via Apify integrations. Keep your other filters (task, library, tags) stable between scheduled runs so the delta stays meaningful.

Input

The simplest input is no input at all — you get the top 100 trending models:

{}

Trending text-generation models built with transformers, top 50:

{
  "dataType": "models",
  "sort": "trending",
  "pipelineTag": "text-generation",
  "library": "transformers",
  "limit": 50
}

Monitor new trending datasets every day:

{
  "dataType": "datasets",
  "mode": "monitor",
  "sort": "trending",
  "limit": 100
}

Most-downloaded text-to-image models this run, models and datasets together:

{
  "dataType": "both",
  "sort": "downloads",
  "pipelineTag": "text-to-image",
  "limit": 100
}

Input	Description
dataType	models (default), datasets, or both
mode	snapshot (default) or monitor (delta — new entries only)
sort	trending (default), downloads, likes, lastModified, createdAt
direction	-1 descending (default, top first) or 1 ascending
pipelineTag	Task filter for models, e.g. text-generation (optional)
library	Library filter for models, e.g. diffusers (optional)
search	Free-text search over names, e.g. "llama" (optional)
tags	Restrict to Hub tags, e.g. ["multilingual"] (optional)
limit	Max results per data type (default 100, up to 1000; paged automatically)
includeConfig	Attach each model's config.json (slower — one extra request per model)
hfToken	Optional Hugging Face token for higher rate limits / gated repos

Output examples

Trending model:

{
  "resultType": "model",
  "modelId": "meta-llama/Llama-3.3-70B-Instruct",
  "author": "meta-llama",
  "modelName": "Llama-3.3-70B-Instruct",
  "pipelineTag": "text-generation",
  "libraryName": "transformers",
  "tags": ["text-generation", "transformers", "conversational", "llama"],
  "downloads": 1840320,
  "likes": 9421,
  "trendingScore": 152.4,
  "lastModified": "2026-06-15T09:31:00.000Z",
  "createdAt": "2026-05-30T09:12:00.000Z",
  "isPrivate": false,
  "gated": "manual",
  "cardData": {
    "license": "llama3.3",
    "language": ["en"],
    "datasets": null,
    "baseModel": ["meta-llama/Llama-3.1-70B"]
  },
  "siblings": 34,
  "modelUrl": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct",
  "trendRank": 1,
  "isNew": null,
  "scrapedAt": "2026-06-19T10:00:00.000Z"
}

Trending dataset:

{
  "resultType": "dataset",
  "datasetId": "HuggingFaceFW/fineweb",
  "author": "HuggingFaceFW",
  "datasetName": "fineweb",
  "tags": ["task_categories:text-generation", "size_categories:10B<n<100B", "language:en"],
  "downloads": 512094,
  "likes": 2103,
  "trendingScore": 88.1,
  "lastModified": "2026-06-01T00:00:00.000Z",
  "createdAt": "2026-04-01T00:00:00.000Z",
  "isPrivate": false,
  "gated": false,
  "cardData": {
    "license": "odc-by",
    "language": ["en"],
    "size": "10B<n<100B",
    "taskCategories": ["text-generation"]
  },
  "datasetUrl": "https://huggingface.co/datasets/HuggingFaceFW/fineweb",
  "trendRank": 1,
  "isNew": null,
  "scrapedAt": "2026-06-19T10:00:00.000Z"
}

Use cases

• ML engineers tracking which models are gaining real adoption before committing to one.
• DevRel teams monitoring ecosystem trends for their framework or library.
• AI researchers spotting emerging architectures, techniques, and base models.
• LLM-ops teams tracking model popularity to inform deployment and support decisions.
• Data teams discovering trending training and evaluation datasets.
• AI newsletters, educators, and content creators covering what's hot on Hugging Face each week.
• RAG pipeline builders tracking models and datasets worth ingesting.

Limitations

• trending is Hugging Face's own signal, not just raw downloads — it blends recency and velocity. Use sort: "downloads" or "likes" if you want a pure popularity ranking.
• Rate limits. The public Hub API is generous but throttles heavy use. The Actor throttles requests and backs off on HTTP 429; increase requestDelayMs, or add an hfToken, if you page large limits.
• Sparse metadata. Not every repo declares a pipelineTag, library, or full card front-matter — those fields come back null rather than failing the run.
• Gated/private repos appear in listings with gated/isPrivate set, but their files aren't accessible without an authorized hfToken.
• includeConfig is slower — it adds one request per model to fetch config.json. Leave it off unless you need architecture details.

Pricing (Pay-Per-Event)

This Actor uses Apify's Pay-Per-Event model — you pay only for what you pull, from ~$4 per 1,000 results:

Event	When charged
Actor run start	Once per run
Trending model tracked	Per trending model returned (in monitor mode, only new-to-trending models)
Trending dataset tracked	Per trending dataset returned (in monitor mode, only new-to-trending datasets)

No subscription, no rental. In monitor mode you're charged only for genuinely new trending entries — checking state is free.

Use it from an AI agent (MCP)

This Actor is MCP-ready: run it as a tool from Claude, Cursor, or ChatGPT via Apify's MCP integration to give your agent live "what's trending on Hugging Face" data — trending models, datasets, and their metadata — on demand.

FAQ

Do I need a Hugging Face API key? No. The public Hub API works without one. Add an optional hfToken only to raise rate limits or read gated/private repos you have access to.

What does "trending" mean vs "most downloaded"? trending uses Hugging Face's own trend signal (recent momentum), while downloads/likes are all-time popularity. Pick the sort that matches your question.

Can I schedule it? Yes — set mode: "monitor" and attach an Apify Schedule (daily/weekly) to get only newly trending entries each run.

Can I filter by task or library? Yes — pipelineTag (e.g. text-generation) and library (e.g. diffusers) for models, plus free-text search and Hub tags for both models and datasets.

How are gated models handled? They're included in the output with gated set ("manual"/"auto") so you can see them trending; downloading their files requires an authorized token.

Does it use a browser? No — it calls the official Hugging Face Hub REST API over HTTP only, which keeps runs fast and cheap.