license: mit
task_categories:
- other
tags:
- huggingface
- spaces
- census
- reproducibility
- link-rot
- software-decay
pretty_name: Hugging Face Spaces Census
size_categories:
- 1M<n<10M
configs:
- config_name: spaces
data_files: spaces.csv
Hugging Face Spaces Census
A complete snapshot of the public Spaces catalogue — 1,458,692 Spaces from 778,898 authors, retrieved 28 August 2026.
This is the population, not a sample. The listing was paginated to exhaustion, with zero duplicate identifiers and no private entries.
The headline
Excluding static Spaces — files on a CDN, which have no runtime and cannot fail — among the 1,150,413 hosted applications:
| Spaces | Share | |
|---|---|---|
| Broken | 443,683 | 38.57% |
| Running | 32,786 | 2.85% |
| Sleeping | 551,048 | 47.90% |
| Paused | 122,663 | 10.66% |
A listed non-static Space is 13.5× more likely to be broken than running.
The platform reports 340,359 Spaces as RUNNING, but 307,573 of those (90.4%) are static pages.
Breakage tracks age
| Created | Spaces | Broken | Rate |
|---|---|---|---|
| 2022 | 16,429 | 14,100 | 85.82% |
| 2023 | 107,879 | 75,974 | 70.43% |
| 2024 | 177,780 | 92,423 | 51.99% |
| 2025 | 653,288 | 151,247 | 23.15% |
| 2026 | 503,316 | 110,127 | 21.88% |
Monotonic across four years, spanning 64 points. Demos that never worked would give a flat rate; a rate climbing with age is dependency drift.
Two definitions that matter
SLEEPING is not broken. A sleeping Space wakes on request and is working software. Counting sleeping Spaces as defects would roughly double the headline and would be wrong. They are reported separately.
Static Spaces are excluded from the denominator. Their observed breakage rate is 0.06% — they have no build step to fail. Including them measures the CDN, not the demos. Both denominators are in the data so either can be recovered.
Files
| File | Contents |
|---|---|
spaces.csv |
One row per Space: id, author, sdk, likes, private, stage, hardware, created_at, last_modified |
census-summary.json |
Precomputed aggregates: stage and SDK distributions, breakage by year, non-static splits |
Validation
Two checks were run against the live API rather than the listing:
- Listing fidelity — 300 random Spaces re-queried individually; 299/300 agreed. The single difference was
RUNNING→SLEEPING, an ordinary idle transition. No error state was misreported. - Failure persistence — 300 Spaces in error states re-queried with retries; 300/300 still broken, none recovered. These are not transient crashes.
Caveats
Nothing here is hidden by the platform. Hugging Face publishes the runtime stage of every Space through its own API. The contribution is the aggregate, which does not appear to have been reported before.
This is a single point-in-time snapshot. The age gradient is cross-sectional — older cohorts are more broken — which supports decay but is not the same as observing individual Spaces break over time.
State is API-reported, not HTTP-probed. A Space marked RUNNING was not confirmed to serve a response, so 2.85% is an upper bound on what actually works.
PAUSED is ambiguous — deliberate author action or exhausted quota — so it is reported separately and excluded from both counts.
Other measures
- 93.18% of Spaces have zero likes; the top 0.1% hold 65.13% of all 1,102,567 likes.
- Breakage by SDK: gradio 44.28%, streamlit 43.50%, docker 28.01%, static 0.06%, no SDK declared 96.09%.
- Authorship is not concentrated — the ten largest authors hold 0.47% of the catalogue, and 66.61% of authors have exactly one Space. Included because the opposite was expected.
Reproducing
Collection and analysis code: github.com/ashishsinha1602/dataset-integrity-audit (paper3/).
Citation
@misc{sinha2026spacescensus,
author = {Sinha, Ashish},
title = {Hugging Face Spaces Census},
year = {2026},
url = {https://huggingface.co/datasets/Ashsinha1/hf-spaces-census}
}
MIT licensed.