Summary
- Only one hard number governs a single write, and the page prints it as prose rather than as a table cell: 500GB per file. The recommendation table on the same page asks for under 200GB per file and leaves the repository-size row without a number, because the page states there is no repository limit in the first place.
- Account allowances run from 100GB of private storage on a free account to 200TB of public base storage on Enterprise, with a 500TB Enterprise Plus row that appears on a second page and nowhere else — while every usage readout I could read is per repository, in bytes, with no reviewed page stating whether those bytes are the file sizes a user uploaded or the deduplicated blocks a content-addressed store actually keeps.
The wall, in three widths
Start with the only hard ceiling in the corpus, printed as a sentence rather than as a cell:
In all cases, no single file will exceed 500GB. I.e. 500GB is the hard limit for a single file size.
The scope is one file. The unit is GB, stated without qualification. And the page does not stop there, because the number a user is more likely to meet is smaller and softer: the same document's repository table recommends a file size of <200GB, fewer than 100k files per repository, fewer than 10k entries per folder and commits of fewer than 100 files (Storage limits). The repository-size cell of that table carries no number at all.
The distinction between those two classes is real, and it lives in prose rather than in the table's markup: one is a recommendation, the other is described by the page itself as a hard limit. What the page does not print is the consequence. The refusal string exists in the corpus, but it arrives from a user rather than from the platform — a forum thread dated 1 May 2025 transcribes 403 Forbidden: Private repository storage limit reached, please upgrade your plan... (Hugging Face Forums). That is a written-down error message, not provider policy, and the same thread's author reports a quota display around 66GB used of 100GB while pushes were blocked. I cite it for one narrow purpose: it is the only text I found where a refusal is written down at all.
The second width is the repository, and here the publisher declines a number deliberately:
There is no per-repo size limit for models and datasets, but uploads count against your account’s total storage quota.
That sentence is the hinge of the whole document. It removes a ceiling at one scope and moves the accounting to another: the repository is unlimited, the account is not. Everything a user actually does — push a model, version a dataset, keep a Space's build artefacts — therefore lands on an allowance denominated in terabytes.
The third width is the account. Free users and organisations get public storage printed as Best-effort* and private storage of 100GB. PRO carries Up to 10TB included* + add-on in public and 1TB + pay-as-you-go private. Team organisations get 12TB base + 1TB per seat + add-on public and 1TB per seat + pay-as-you-go private. Enterprise organisations get 200TB base + 1TB per seat + add-on public, with a separate Up to 1,000TB for large contracts, and 1TB per seat + pay-as-you-go private. Public storage is also sold by the terabyte: 1TB at $12 per month, 5TB at $60, 10TB at $120, 20TB at $240, 50TB at $500 — the last row working out to $10 per terabyte per month, the only place on the page where buying more costs less per unit. Private pay-as-you-go starts at $18 per TB per month and steps down to $16 at 50TB, $14 at 200TB and $12 at 500TB.
Free public storage is printed as Best-effort*, with mitigations described as preventing abuse of free public storage. It is not a numeric ceiling and I am not going to invent one for it.
Two pages, one quantity, three hundred terabytes apart
The Enterprise row is where the corpus splits. The storage-limits table stops at Enterprise with 200TB of public base storage and carries no Enterprise Plus row at all. The plan comparison on the enterprise page prints four columns: Free Best effort, Team 12TB base + 1TB/seat, Enterprise 200TB base + 1TB/seat, and Enterprise Plus 500TB base + 1TB/seat (Team & Enterprise plans). Enterprise Plus, at 500TB of public base storage, appears nowhere else in the ten documents.
Two provider pages, two different top rows, no cross-reference between them and no visible date on either. The honest description is narrow: one page lacks a row the other carries. I will not call the first table stale, and neither page gives me the means to date them against each other.
The private-storage overage splits the same way. The billing page prints one mechanic — Above the included 1TB (or 1TB per seat) of private storage in PRO, Team, and Enterprise, additional private storage is billed in 1TB increments, at a base price of $18/TB/month. (Billing) — while the storage-limits page prints continuous pay-as-you-go with volume steps. Increments at a base price and a rate with stepped tiers describe the same meter two ways, and neither page names the other.
What counts
The scope sentence is broad and names its own exceptions: Storage limits and policies apply to all types of repositories (models, datasets, buckets, …) on the Hub. Then the limitations section carves buckets out with This section does not apply to Storage Buckets, so the recommendation class and the storage-bucket class are documented as separate things on one page.
What that leaves is version history, large-file objects, dataset files and repository contents, all counted at the account. The storage backend explains why the count is not self-evident. Xet, the Hub's content-addressed store, keeps large files as pointer files with .gitattributes metadata while the actual data is stored in remote storage (like Amazon S3), applies chunk-level deduplication, and leaves Git LFS supported for compatibility (Xet). The same page volunteers: Large - model or dataset files are in the range of GB and above. We have a few TB-scale files!
That sentence and the 500GB hard cap sit in one corpus without a reconciling line. I am not going to declare the TB-scale files exempt, and I am not going to declare them in breach. The documents state the cap, observe the files, and leave the reader holding both.
A Space's disk is a different object again, and it is the one place a small numeric disk figure is printed: Each Spaces environment is limited to 16GB RAM, 2 CPU cores and 50GB of (not persistent) disk space by default, which you can use free of charge. (Spaces Overview) The accounting here is per environment, and the parenthetical matters more than the number: non-persistent disk is not storage a user keeps. The Spaces storage page prints no gigabyte figure at all; it describes a small amount of disk that is ephemeral and routes persistence to Storage Buckets mounted as volumes.
The viewer's 5GB is the trap in this corpus. It governs what the Dataset Viewer can index, sort and search — the full dataset is shown for Parquet with sorting and search only on the first 5GB, non-Parquet data above 5GB shows a partial viewer, and the first 5GB of every dataset is auto-converted to Parquet (Data Studio). It is a scope for reading, not a quantity a user may store, and anyone quoting it as a storage ceiling has misread the page.
The ruler
Here is what the account holder is actually given to measure against those ceilings.
From 18 September 2025, the Files & Versions tab displays the total size of a repository on the current branch/folder, updating as files change (Changelog). That is a per-repository number, and the scope is the branch or folder — not the repository's accumulated history.
A settings view followed: User and organization settings now include a Repositories page to visualize repository storage consumption (Changelog). The entry frames the page as monitoring usage, understanding how storage is distributed across repositories, and managing resources; it prints no ceiling and no percentage of one.
Programmatically, list_user_repos() returns repositories for an account or organisation — models, datasets, Spaces and buckets — and Results include storage information and are sorted by storage usage. The CLI equivalent is hf repos ls, and the documented example output prints a byte count beside a repository name (huggingface_hub).
Three surfaces, one scope: per repository, in bytes. Even the storage-limits page's remedy for running out is per repository — Settings, a List LFS files control, and a note that closing or merging pull-request refs shows the estimated space that could be freed. No document among the ten prints an account-level remaining allowance as a number. The billing page does say You can monitor your usage at any time from your billing dashboard, located in your user's or organization's settings menu, but that sentence sits in the compute-services section, and the retained copy does not state whether the dashboard covers storage.
So the reading position is this. A user facing a 200TB public allowance and a 1TB private inclusion can see, per repository, a byte total. To compare it with the allowance they must add repositories themselves — and then decide what the resulting number means.
The unit nobody printed
A content-addressed store with chunk-level deduplication keeps two quantities that need not be equal: the logical size of the files a repository presents, and the physical blocks the store retains for them. A ceiling in terabytes and a usage figure in bytes describe the same object only if both count the same bytes.
In the ten documents I read, no page states which one the usage readouts count. The hub guide does not say. The Xet page describes deduplication and never connects it to quota accounting. The storage-limits page prints ceilings in GB and never a byte count beside one. The one place a byte count appears is the documented example output — 4,828,692,480 bytes beside a repository name — which is also the only place in the corpus where the ruler's unit is explicit.
My arithmetic, labelled mine
Every figure below is my own division of numbers the documents print. None of it is the provider's, and none of it settles a question the documents leave open.
500GB × 100,000 files is 50,000,000 GB, or 50 PB, for a single repository configured to the top of the recommendation table's file-size and file-count rows. 50 PB ÷ 10TB = 5,000: the PRO public allowance is one five-thousandth of that repository. On the recommendation row rather than the hard cap, 200GB × 100,000 = 20 PB, and 20 PB ÷ 10TB = 2,000. 500GB ÷ 100GB = 5, so one file at the hard cap is five times the entire free private allowance. 10TB ÷ 500GB = 20 files at the hard cap fit inside PRO's public inclusion; 200TB ÷ 500GB = 400 fit inside Enterprise's public base.
200GB ÷ 5GB = 40: the viewer's indexing scope is one fortieth of the recommended single-file size. 500GB ÷ 50GB = 10: one permitted file is ten times a Space's default non-persistent disk. And 4,828,692,480 ÷ 1,000,000,000 = 4.83 decimal GB, while ÷ 1,073,741,824 = 4.50 binary GiB — a gap of about seven percent between two readings of one byte count, which is the size of the ambiguity nothing in the corpus resolves.
Where this reading stops
The pricing page returned only decorative glyph rows in the retained copy and a follow-up search came back empty, so nothing here is attributed to it and I do not characterise it. The models-uploading page returned no readable text. The repositories-recommendations address returned a body identical to the storage-limits page, so it is treated as a duplicate and cited nowhere. The Spaces storage page prints no gigabyte figure. Older mirrors of the upload guidance print a 5GB recommendation and a 50GB per-file hard limit; those are historical doc versions, cited nowhere and used for nothing.
What remains is a platform that did the harder editorial work — a hard limit separated from a recommendation in prose, an explicit refusal to pretend a repository ceiling exists, non-persistent disk labelled as such — and has not printed the one document its own allowance table implies: a page where a user sees the measured quantity standing against the ceiling, in the same unit, with the unit defined.
The subject has its own entry in our directory: Hugging Face.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






