Model Pulse

Daily download history for every model and dataset on the Hugging Face Hub, and likes for every Space, back to July 2024.

A model page on the Hub tells you its downloads for the last 30 days, but not how it got there. Model Pulse rebuilds the day-by-day history for 1.6M models, 1.06M datasets and 104K Spaces from the daily snapshots in cfahlgren1/hub-stats, and keeps it updated every day.

• Site: modelpulse.ifsp.dev, also as a Hugging Face Space
• Open data: modelpulse/model-pulse-data, one row per model per day
• Report: What 19 months of daily downloads say about the Hub

What you can do with it

• Look up any model, e.g. modelpulse.ifsp.dev/model/Qwen/Qwen3-8B: weekly, daily, 30-day and all-time downloads, likes, ranks and milestones

• Look up any dataset or Space, e.g. HuggingFaceFW/fineweb: downloads or likes over time, the models trained on a dataset, and the Spaces using a model or dataset

• Compare up to five models on one chart

• See a model's family: how much of its reach comes from its quantizations, fine-tunes, adapters and merges

• Galaxy: every model built on a base model, drawn as one picture

• Wrapped: any author's last 12 months on the Hub in six shareable cards

• Rankings: most downloaded, fastest growing, new breakouts, biggest families, top organizations

• A badge for your model card, refreshed daily:

  

How it works

1. Backfill. pipeline/backfill_xet.py reads every daily revision of hub-stats models.parquet and keeps four columns per model per day. pipeline/build.py turns them into monthly partitions, per-model metrics, family trees (via base_model links), author and Hub-wide series, and the leaderboards.
2. Daily update. pipeline/daily.py appends each new snapshot, recomputes the metrics and uploads the changed files to the dataset. It runs inside the API container, which checks for new data every hour.
3. API and site. space/app is a FastAPI + DuckDB service over the dataset (JSON API, badges, server-rendered pages, link previews and sitemaps). space/web is the Preact + uPlot frontend, served both on its own domain and as a static Hugging Face Space.

Download counts follow the Hub's own counting rules. All-time totals start on 2025-02-27, when the Hub began reporting them; the rolling 30-day series goes back to 2024-07-29.

Repository layout

Path	What
pipeline/	Backfill, full build and daily incremental update of the dataset (polars): build.py/daily.py for models, build_repos.py/daily_repos.py for datasets and Spaces, stalls.py for days when the Hub's counters stood still (fully or in part) or went backwards, and rebuild_hub.py to rebuild the Hub-wide series and skip days from the snapshots
space/app/	FastAPI + DuckDB API, badges, server-rendered pages (seo.py), link previews (og.py), datasets and Spaces (repos.py), background jobs
space/web/	Preact + Vite + uPlot frontend
brand/	Logo, thumbnail and organization card

Running it

# frontend, against the public API
cd space/web && npm install && VITE_API_BASE=https://modelpulse.ifsp.dev npm run dev

# API, with the dataset downloaded to ./data
cd space && pip install -r requirements.txt
DATA_DIR=../data DATA_REPO=modelpulse/model-pulse-data python -m app.serve

space/deploy.sh builds the frontend and runs the API container on a server behind Traefik (set DEPLOY_HOST in space/.deploy.env). publish_site.sh publishes the frontend to the Space.

Credits

Made by Tardelli Stekel (@tardellirs, stekel.ifsp.dev). Built on the daily snapshots collected by @cfahlgren1. An independent project, not affiliated with Hugging Face.

Licensed under the Apache License 2.0.