modelpulse/model-pulse-data

Website · Space · Quick start · Tables · Methodology · Citation

Model Pulse data

A daily time series of the whole Hugging Face Hub. Downloads and likes for every actively used model and dataset, and likes for every liked Space, one row per repo per day since 2024-07-29, refreshed every day.

It is the data behind Model Pulse, which has a page with the full history of every model, dataset and Space, along with model families (quantizations, fine-tunes, adapters and merges), author totals and weekly rankings.

Coverage	2024-07-29 → today, one snapshot per day
Models	1.61M tracked · 25K families · 700M daily rows
Datasets	1.06M tracked · 365M daily rows
Spaces	104K tracked (at least one like) · 45M daily rows
Total	~1.15B rows · ~2 GB of Parquet
Update	Daily, shortly after each new source snapshot
License	Apache 2.0

Quick start

Every table is plain Parquet, so it can be queried in place without downloading the whole repo.

DuckDB: daily downloads of one model

SELECT day, dl30, dl_all - lag(dl_all) OVER (ORDER BY day) AS downloads
FROM 'hf://datasets/modelpulse/model-pulse-data/series/*.parquet'
WHERE id = 'Qwen/Qwen3-8B'
ORDER BY day;

Polars: this week's most downloaded models

import polars as pl

top = (
    pl.read_parquet(
        "hf://datasets/modelpulse/model-pulse-data/models.parquet",
        columns=["id", "pipeline_tag", "dl_7d", "growth_7d", "rank_7d"],
    )
    .filter(pl.col("rank_7d").is_not_null())
    .sort("rank_7d")
    .head(20)
)

🤗 Datasets: load any table by its config name

from datasets import load_dataset

hub = load_dataset("modelpulse/model-pulse-data", "hub_series", split="train")
models = load_dataset("modelpulse/model-pulse-data", "models", split="train")

The daily series are split into monthly files (series/2026-09.parquet, …). To work with a recent window, read only the months you need.

Tables

Each table below can be loaded as a config.

Config	Path	One row per	Rows
series	series/YYYY-MM.parquet	model × day	700M
models	models.parquet	model	1.61M
family_series	family_series/YYYY-MM.parquet	base model × day	12M
author_series	author_series/YYYY-MM.parquet	author × day	25M
hub_series	hub_series.parquet	day × task (pipeline_tag)	32K
dataset_series	datasets/series/YYYY-MM.parquet	dataset × day	365M
datasets	datasets/datasets.parquet	dataset	1.06M
space_series	spaces/series/YYYY-MM.parquet	Space × day	45M
spaces	spaces/spaces.parquet	Space with at least one like	104K
uses	uses.parquet	reference (Space → model/dataset, model → dataset)	464K

Daily series: series, dataset_series, space_series, family_series, author_series, hub_series

Column	Type	Description
id	string	Repo id (org/name), or the base model for family_series
author	string	Author or organization (author_series only)
day	date	Snapshot day
dl30	int	The Hub's rolling 30-day download count
dl_all	int	All-time downloads (available from 2025-02-27)
likes	int	Likes on that day
trending	float	Trending score (space_series)
members / models	int	Repos in the family / models by the author on that day
pipeline_tag, dl	string, int	Task and Hub-wide downloads that day (hub_series, from 2025-02-28)

In family_series and author_series, downloads are summed over the base model and all its derivatives, or over all of the author's models.

Models: models.parquet

Column	Description
id, author, pipeline_tag, library_name, license, params, is_gguf	Latest metadata from the Hub
created_at, last_modified, first_seen	Creation and update times, and the first day the model appears in the data
dl30, dl_all, likes, trending	Latest counters
dl_7d, dl_prev7d, dl_base7d, growth_7d, likes_7d, peak_share, steps	Weekly figures (see Weekly figures)
rank_dl30, rank_task, rank_7d	Overall rank by 30-day downloads, rank within the task, rank by 7-day downloads
base_relation, base_ids	How the model derives from its base model(s): finetune, adapter, quantized or merge
fam_members, n_quantized, n_finetune, n_adapter, n_merge	Size of the family built on this model
fam_dl30, fam_all, fam_7d (and *_desc)	Downloads summed over the model and its family (*_desc: derivatives only)

Datasets: datasets/datasets.parquet

The same counters, weekly figures and ranks as for models, plus gated, description, pipeline_tag, size, license and modality, and the usage counts used_by_models and used_by_spaces.

Spaces: spaces/spaces.parquet

Column	Description
id, author, title, emoji, short_description, sdk	Latest metadata
created_at, last_modified, first_seen	Creation and update times, first day in the data
likes, likes_7d, likes_30d, trending	Likes in total and gained over 7 and 30 days, trending score
rank_likes, rank_7d	Rank by total likes and by likes gained this week
uses	Number of models and datasets the Space references

The Hub doesn't publish visit counts for Spaces, so Spaces are tracked by likes.

Usage graph: uses.parquet

Column	Description
src_kind, src	The Space or model that references something
dst_kind, dst	The model or dataset it references
created	Creation date of the source repo
weight	Popularity of the source: its downloads for a model, its likes for a Space

These rows come from the cards as they are today, dated by when each Space or model was created, not by when it started using what it lists.

Other files

Path	Contents
children.parquet	One row per base → derivative edge: parent, id, base_relation, author, dl30, dl_all, dl_7d, likes
renames.parquet, datasets/renames.parquet	Repos renamed on the Hub: old, new, gone, came (see Renames)
spaces/new_by_sdk.parquet	Spaces created each day by SDK, from the latest snapshot
leaderboards.json, datasets/leaderboards.json, spaces/leaderboards.json	The weekly rankings shown on the site
meta.json (models), repos_meta.json (datasets and Spaces)	Days covered, build and snapshot times, and the day lists used for quality control: skip_days, low_days, short_days, partial, frozen

Methodology

The data is rebuilt from the daily snapshots of cfahlgren1/hub-stats by reading every historical revision of its models.parquet, datasets.parquet and spaces.parquet. The Hub's counters are noisy in a few known ways, and the rules below make the daily series reliable without changing any totals.

Tracking

• A model or dataset is tracked once it has 10+ downloads in 30 days, 50+ all-time downloads, or at least one like.
• A Space is tracked once it has at least one like.
• Download counts follow the Hub's download counting rules. The Hub occasionally books delayed downloads on a single day.

Daily downloads

• dl30 is the Hub's rolling 30-day count. dl_all (all-time downloads) only exists from 2025-02-27, so exact daily downloads, computed as the difference of dl_all between snapshots, start on that date.
• When a repo's all-time counter hasn't moved in 31 days, dl30 is set to 0, because the Hub sometimes keeps showing a stale 30-day count.

Missing and partial snapshots

• Some days are missing from the source (Aug 2024, Jun 2025, Apr 2026, May–Jun 2026). Totals are unaffected: the days without a snapshot share the per-day average of the next snapshot.
• A few snapshots hold only part of the repos (for models, 2025-11-24 has 289K of 1.13M). A snapshot with fewer than half the rows of the one before is left out: nothing is measured from or to it, the gap around it is spread like any missing day, and it is listed in partial and in skip_days.

Counter stalls

On some days the Hub's download counters stand still and catch up a day or two later, mostly on Wednesdays. Dataset counters freeze all at once. Model counters often freeze only in part (newer repos stop while older ones keep counting), so the model total may only drop by half.

A day is treated as a stall when any of these is true:

• the total is under 0.3 of its 15-day median;
• at least 25% of the models with 60K+ monthly downloads didn't move at all;
• a dip under 0.7 is made up by the next or the previous day. Here, a value spread over days without a snapshot counts as one measurement, with its days weighed in the sum.

hub_series spreads each stall evenly over its days, so sums don't change. The snapshots inside a stall are listed in skip_days so that per-repo series can be smoothed the same way, and frozen holds the recent daily share of frozen counters. A stall that reaches into a stretch without snapshots takes the whole stretch, up to the snapshot that closes it, so a catch-up booked across missing days is spread together with the stall it belongs to.

Counter regressions

Twice (May 2025 and June 2026), the all-time counters went down for a large share of models and came back days later. Those snapshots are also in skip_days. Downloads across them are measured from the last good snapshot to the first one after the counters recovered, so the recovery isn't counted as new downloads. A drop that doesn't recover within 8 snapshots is treated as a lasting correction by the Hub and is not set aside.

Low and short days

• low_days lists the few days that are still well below their local median after all the rules above, and that the days around them don't make up for. A day is low when it is under 0.7 of the 15-day median (adjusted for the day of the week and estimated from days with their own snapshot, since weekends run a little lower), in a run that the 3 days on each side don't make up by at least half. Days without a snapshot share one measurement with the snapshot that closes them, so they are judged together. For example, 2026-01-31 (no snapshot) and 2026-02-01 come out at 0.74 as one measurement, so they are not low. Low days keep their measured values, and the Hub chart greys them out instead of smoothing over them. The last two days are only judged once their neighbours are known.
• short_days lists the snapshots taken less than 18 hours after the previous one. This happened when hub-stats moved its collection time (2025-03-04 came 10.3 hours after 2025-03-03). The counters' daily update then lands in the next snapshot, so such a day is short, not low, and never counts as a low day. snapshot_at keeps the time of the latest snapshot.

Weekly figures

• dl_7d, likes_7d and growth_7d are rates over the days that the reference snapshots actually span, so a skipped or missing snapshot never makes a week longer than 7 days.
• A repo with no reference snapshot has no weekly figure, unless it is new that week.
• The growth rankings leave out repos whose week came mostly from a single day: over 60%, or over 80% for repos created that month once they have three snapshots. peak_share holds that share.

Renames

A repo counts as renamed when its old id disappears and a new id appears within 3 days, under the same name or the same owner, with the same all-time count (within 1%). On the site, old pages redirect to the new id, whose series and author and family totals include the old id's history.

Source and license

• Source: cfahlgren1/hub-stats (Apache 2.0), daily snapshots of the Hub API.
• License: Apache 2.0.
• Model Pulse is an independent project and is not affiliated with Hugging Face.

Citation

@misc{stekel2026modelpulse,
  author       = {Stekel, Tardelli},
  title        = {Model Pulse: Daily History of Hugging Face Models, Datasets and Spaces},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/modelpulse/model-pulse-data}}
}

Contact

Made by Tardelli Stekel (@tardellirs · stekel.ifsp.dev). Questions, corrections and ideas are welcome in the Community tab.