fshy / Compute / Debrid / Inference / Storage

Hugging Face

huggingface.co · Compute, Debrid, Inference, Storage · 13 records · JSON

Hugging Face Hub Debrid★ qualifiesanonymousno cardMper-ip-shaping
Hugging Face Community Computecaveatemailno cardDwrite-gated
Hugging Face Spaces Computecaveatemailno cardDretention-clampeddisagrees
Hugging Face ZeroGPU Computecaveatemailno cardDvolume-quota
Hugging Face Storage Buckets via S3 gateway (s3.hf.co) — as an annex/DVC backend Storagecaveatinstitution?card?Rper-object-cap
Hugging Face Inference Computepaidemailno cardD
Hugging Face Jobs (hf-sandbox backend) Computepaidemail?card?Drequest-rate
Hugging Face Spaces (CPU Basic) ComputepaidcardcardDwrite-gateddisagrees
Hugging Face Spaces Dev Mode ComputepaidcardcardR
Hugging Face Inference Providers InferencepaidcardcardU
Hugging Face Spaces (disk check only) Storagefails: quotaemail?card?R
hugging face spaces disk (as persistent storage) Storagefails: quotaemail?card?R
hugging face storage buckets (re-check only — not a discovery) Storagefails: quotaemail?card?R

Hugging Face Hub Debrid

★ qualifiesanonymousno cardMper-ip-shaping

huggingface.co/docs/hub/rate-limits · model and dataset registry with a CDN (anonymous read) · debrid/hugging-face-hub · JSON · compare

Free tier	Anonymous download of any public model or dataset file, no token. The published per-5-minute windows are API 500, Resolvers 3,000, Pages 100 for an anonymous caller, and 1,000 / 5,000 / 200 for a free account; 'All quotas are calculated over 5-minute fixed windows.' unit requests; metered on request-count; Bytes are not metered at all; only request counts are.
Price	USD 0
Per-object cap	not applicable per file; a repository may hold very large files. The file measured here is 548,118,077 bytes (openai-community/gpt2 pytorch_model.bin).
Retention	until the repository owner deletes or revises the revision; anonymous readers can pin a commit hash to keep a fixed object
Egress	Free and unmetered. /resolve/ redirects to a CDN host (us.aws.cdn.hf.co through the xet bridge) with a short-lived signed URL; the signature is issued to the anonymous caller and no token is involved. free_allowance: no byte allowance published
Identity	anonymous No token: https://huggingface.co/api/whoami-v2 returned 401 for an anonymous client, and the same client downloaded 548,118,077 bytes from /resolve/ successfully. A free account doubles the request windows; it buys no bytes.
Limits	request-rate anonymous: 500 API, 3,000 Resolvers, 100 Pages per 5-minute window; free account doubles all three source per-ip-shaping measured here: one connection sustained 34.8-36.6 MiB/s on a 256 MiB range and 39.98 MiB/s on a 548 MB file, while eight parallel range requests totalled only 17.1-21.3 MiB/s Whether the 8-connection drop is Hugging Face's CDN, the ISP shaping parallel flows, or the line itself was not established: the Cloudflare speed test endpoint refused for a user with 403, so there is no independent line-ceiling measurement.
Binding	per-ip-shaping Request counts never bind at this scale - 3,000 resolver calls per 5 minutes is 864,000 a day, far more files than a TiB holds. What limits a day is the ~36-40 MiB/s ceiling, and adding connections lowers it rather than raising it.
Interface	rest, cli, git GET https://huggingface.co/<repo>/resolve/<rev>/<path> for a file; the API at /api/models and /api/datasets; the dataset viewer at datasets-server.huggingface.co (rows, parquet, size); hf CLI and git for repo-level sync. endpoint: https://huggingface.co/api/models
Tooling	hf CLI, huggingface_hub python, git-lfs, plain curl on /resolve/. rclone reads it through the HTTP remote. The dataset viewer's /parquet endpoint hands back per-file parquet URLs that a client can fetch directly.
Composes with	The fastest measured pipe and the only one whose content is current ML data rather than a fixed corpus: pull a dataset's parquet parts for training, and mirror the same files to storage with rclone. Common Crawl now mirrors selected crawls into an HF Storage Bucket, so both paths can share one client.
Reliability	The /resolve/ URL redirects to a signed CDN URL that expires; a client must follow the redirect per request rather than cache the final URL. Anonymous windows are explicitly 'subject to change over time depending on platform health'.
Notes	Answer sits here: about 40 MiB/s over one connection with no account, which is 3.3 TiB/day if the rate held for a day. The eight-connection run is the trap - it is half the speed, so a client that opens eight streams to go faster gets there slower.
Evidence rung	measured A complete 548 MB file was downloaded and hashed, and three 256 MiB runs at each concurrency were recorded. The extrapolation to bytes/day is arithmetic on those numbers, not a 24-hour run.
Checked	2026-10-24 (live); stale on 2026-11-23
Parsed numbers	`$.limits[1].value` limit:per-ip-shaping 36.6 MiB `$.price_per_tb_month` price-per-tb USD0.0 TB / month

Evidence (6)

date	source	what it proved
2026-10-23	huggingface.co/docs/hub/rate-limits	the per-5-minute windows: anonymous 500 API / 3,000 Resolvers / 100 Pages; free account 1,000 / 5,000 / 200
2026-10-23	probe:https://huggingface.co/openai-community/gpt2/resolve/main/pytorch_model.bin	full file: 200, 548,118,077 bytes, sha256 7c5d3f4b8b76583b422fcb9189ad6c89d5d97a094541ce8932dce3ecabde1421, 13,072.6 ms = 39.98 MiB/s over one connection
2026-10-23	probe:https://huggingface.co/openai-community/gpt2/resolve/main/pytorch_model.bin --range 0-268435455	256 MiB at one connection: 36.593, 34.818, 36.610 MiB/s across three runs
2026-10-23	probe:https://huggingface.co/openai-community/gpt2/resolve/main/pytorch_model.bin --concurrency 8 --shard-bytes 268435456	256 MiB over eight connections: 17.058, 21.332, 20.440 MiB/s total, each below its own one-connection baseline - parallelism loses
2026-10-23	probe:https://huggingface.co/api/whoami-v2	401 with www-authenticate: Bearer - there was no token in the client that downloaded the 548 MB file
2026-10-23	probe:https://datasets-server.huggingface.co/rows?dataset=openai/gsm8k&config=main&split=train&offset=0&length=100	200, 61,691 bytes for 100 rows, anonymous - the dataset viewer is readable without a token

More fields (6)

{
 "full_file_sha256": "7c5d3f4b8b76583b422fcb9189ad6c89d5d97a094541ce8932dce3ecabde1421",
 "full_file_bytes": 548118077,
 "one_conn_mib_s": [
  36.593,
  34.818,
  36.61
 ],
 "eight_conn_mib_s": [
  17.058,
  21.332,
  20.44
 ],
 "full_file_mib_s": 39.98,
 "arithmetic_tib_per_day_at_40_mib_s": 3.29
}

Hugging Face Community Compute

caveatemailno cardDwrite-gated

huggingface.co/docs/hub/spaces-gpus · GPU cloud · compute/hugging-face-community · JSON · compare

Free tier	Quoted: 'Do you have an awesome Space but need help covering the GPU hardware upgrade costs? We love helping out those with an innovative Space so please feel free to apply for a community GPU grant and see if yours will make the cut! This application can be found in your Space hardware repo settings in the lower left corner under "sleep time settings"'. There is also a lighter path, quoted: 'You can even request a free upgrade if you are building a cool demo for a side project!' The unit is a hardware tier attached to one Space, not a credit balance, and billing is quoted as 'based on hardware usage and is computed by the minute'. unit hours; metered on wall-clock; grant
Price	none stated
Per-object cap	not applicable
Retention	not applicable
Egress	not stated
Identity	email not published
Limits	write-gated the grant is attached to a single Space rather than to an account: approval is discretionary, it covers the hardware tier of that Space only, and nothing in the record says the hours can be moved to a training job. source the grant is attached to a single Space rather than to an account: approval is discretionary, it covers the hardware tier of that Space only, and nothing in the record says the hours can be moved to a training job.
Binding	write-gated the grant is attached to a single Space rather than to an account: approval is discretionary, it covers the hardware tier of that Space only, and nothing in the record says the hours can be moved to a training job.
Interface	rest inferred from the tooling the provider documents; not probed
Tooling	not stated
Composes with	The grant is the only route found in this corpus to a dedicated, always-on GPU behind a public URL with no card: a free account gets 2 ZeroGPU Spaces with 5 minutes of GPU a day (dossier compute record huggingface-spaces-zerogpu), and this grant is what a project applies for when 96 GB and a 5 minute window are not enough. It composes with the corpus's inference records because the Space can sit in front of an external API and hold the connection.
Reliability	no reliability event recorded
Notes	Distinct from the ZeroGPU record already in the corpus: ZeroGPU is a quota every free account holds, while a community GPU grant is a discretionary hardware allocation to one Space. The amount is unverified and deliberately so: the docs publish no tier list, no hour count and no approval rate, and the public applications found on Spaces range from a demo to 2,500 GPU hours, so the honest claim is the mechanism. Three readings of 'community grant' were considered: hardware only, hardware plus storage, and a cash-equivalent credit. The docs describe hardware ('covering the GPU hardware upgrade costs') and a public applicant asked for storage separately, which supports hardware-only.
Evidence rung	documented
Checked	2026-10-22 (live); stale on 2026-11-21

Evidence (2)

date	source	what it proved
2026-10-22	huggingface.co/docs/hub/spaces-gpus	'Do you have an awesome Space but need help covering the GPU hardware upgrade costs? We love helping out those with an innovative Space so please feel free to apply for a community GPU grant'; 'This application can be found in your Space hardware repo settings in the lower left corner under "sleep time settings"'; 'You can even request a free upgrade if you are building a cool demo for a side project!'; 'Billing on Spaces is based on hardware usage and is computed by the minute'; page 200
2026-10-22	huggingface.co/spaces/cicero-im/marimo/discussions/1	an open application on a public Space requesting '2,500 GPU hours for model training and evaluation' and '5TB for datasets', which shows the request channel is public and the ask is negotiable rather than fixed (reported: one applicant's request, not an award)

More fields (1)

{
 "wave2": {
  "class": "community-grant",
  "free_tier_kind": "grant",
  "shape": {
   "vcpu": null,
   "ram_gib": null,
   "gpu_class": null,
   "gpu_count": 1
  }
 }
}

Unknown or withheld values (9)

$.egress.metered egress-unknown
$.egress.terms not-stated
$.identity_cost.detail process
$.per_object_cap.notes process
$.per_object_cap.stated not-applicable
$.reliability_note none-observed
$.retention.notes process
$.retention.stated not-applicable
$.tooling not-stated

Hugging Face Spaces Compute

caveatemailno cardDretention-clampeddisagrees

huggingface.co/docs/hub/en/spaces-overview · platform host · compute/huggingface-spaces · JSON · compare

Sibling records disagree:

verdict: compute/huggingface-spaces says qualifies-with-caveat, compute/huggingface-spaces-cpu-basic says paid-only
identity_kind: compute/huggingface-spaces says email, compute/huggingface-spaces-cpu-basic says card

Free tier	Static Spaces are free for everyone; Gradio and Docker Spaces require a paid plan to create, PRO for personal accounts; free accounts in good standing may host up to 2 ZeroGPU Spaces unit none; metered on wall-clock; the CPU Basic hardware still lists as FREE but you cannot create a Space that uses it on a free account
Price	none stated
Per-object cap	not applicable
Retention	a Space on free hardware sleeps after a period of inactivity; a paused Space stops executing until restarted
Egress	not priced on the page read
Identity	email a verified email and an account older than 30 days; no card for the free lane
Limits	write-gated creating a Gradio or Docker Space needs PRO or higher; a free account cannot create one source per-object-cap each Space is limited to 16 GB RAM, 2 CPU cores and 50 GB of non-persistent disk source retention-clamped a Space on free hardware sleeps after a period of inactivity source ZeroGPU on a free account allows 5 minutes of GPU a day and needs an account older than 30 days
Binding	retention-clamped the 5 minutes of GPU a day and the sleep-on-inactivity bind long before the 2-Space limit does
Interface	rest, cli Gradio, Docker or static SDK; ZeroGPU for Gradio on a free account
Tooling	huggingface_hub, the Space SDK, git push to the Space repo
Composes with	ZeroGPU is a GPU allowance of 5 minutes a day, which is an inference budget and not a machine
Reliability	no reliability event recorded
Notes	The free 2 vCPU / 16 GB CPU Basic hardware still exists and still costs nothing; a free account cannot create a Space that runs on it. talaria demoted this row correctly and nemo did not. What survives on a free account is a static Space, which has no runtime. Corrected 2026-10-22: paid-only was wrong. A free lane exists, limited to 2 ZeroGPU Spaces for a personal account with a verified email older than 30 days. The ZeroGPU table prices it at 5 minutes of GPU a day.
Evidence rung	documented
Checked	2026-10-22 (live); stale on 2026-11-21
Parsed numbers	`$.extras.m10h_usd` price USD9.0 / month `$.extras.m24h_usd` price USD9.0 / month `$.limits[1].value` limit:per-object-cap 16.0 GB

Evidence (4)

date	source	what it proved
2026-10-22	huggingface.co/docs/hub/en/spaces-overview	'Static Spaces are free for everyone. Gradio and Docker Spaces run on compute and require a paid plan to create: PRO for personal accounts'
2026-10-22	huggingface.co/docs/hub/en/spaces-overview	free personal accounts in good standing can host up to 2 Gradio Spaces on ZeroGPU
2026-10-22	huggingface.co/docs/hub/en/spaces-overview	each Space is limited to 16 GB RAM, 2 CPU cores and 50 GB of non-persistent disk; CPU Basic lists as FREE hardware
2026-10-22	huggingface.co/docs/hub/spaces-overview	'Static Spaces are free for everyone. Gradio and Docker Spaces run on compute and require a paid plan to create: PRO for personal accounts, Team or Enterprise for organizations. Free personal accounts in good standing can still host up to 2 Gradio Spaces running on ZeroGPU', quoted

More fields (13)

{
 "shape": {
  "vcpu": 2,
  "ram_gib": 16,
  "disk_gib": 50,
  "gpu_class": null,
  "gpu_count": null
 },
 "billing_model": "wall-clock",
 "granularity": {
  "min_billed_s": 60,
  "rounding_s": 60
 },
 "session_cap": {
  "wall_clock_min": null,
  "idle_reap_s": null,
  "restart_required": false,
  "note": "a free-hardware Space sleeps after inactivity; the interval is not stated on this page"
 },
 "keep_rate": {
  "value": 0.0,
  "basis": "a sleeping Space is not billed; CPU Basic has no hourly cost",
  "source": "https://huggingface.co/docs/hub/en/spaces-overview"
 },
 "persistent_storage": {
  "included_gib": 50,
  "price_gib_month": "not persistent",
  "durability": "none",
  "survives_stop": false,
  "survives_destroy": false,
  "max_retention": "the Space lifetime"
 },
 "network_out": {
  "allow_outbound": true,
  "allow_inbound_listen": true,
  "ports": "HTTPS on the Space URL",
  "static_ip": false,
  "ipv6": null,
  "per_session_bytes": null
 },
 "free_tier_kind": "none",
 "m10h_usd": 9.0,
 "m24h_usd": 9.0,
 "price_arithmetic": "USD 9 a month for PRO is the entry price for a compute Space, before hardware upgrades",
 "census_rows": [
  "huggingface-spaces"
 ],
 "wave2": {
  "verdict_change": "paid-only -> qualifies-with-caveat on 2026-10-22"
 }
}

Unknown or withheld values (3)

$.per_object_cap.notes process
$.per_object_cap.stated not-applicable
$.reliability_note none-observed

Corrections

2026-10-22 Hugging Face Spaces has a free lane, and the ZeroGPU allowance recurs daily

Hugging Face ZeroGPU Compute

caveatemailno cardDvolume-quota

huggingface.co/docs/hub/en/spaces-zerogpu · notebook runtime · compute/huggingface-spaces-zerogpu · JSON · compare

Free tier	a free account in good standing (verified email, older than 30 days) may host up to 2 ZeroGPU Spaces and use 5 minutes of GPU a day, on a half RTX Pro 6000 Blackwell with 48 GB VRAM unit minutes; metered on wall-clock; The quota is per day and resets 24 hours after first use. Only the Gradio SDK is compatible.
Price	none stated
Per-object cap	not applicable: the unit is a session, not an object
Retention	the Space repo persists; the GPU allocation is per function call and released on completion
Egress	no egress allowance published
Identity	email a verified email and an account older than 30 days; no card. The 30-day age is the real gate and cannot be bought past.
Limits	volume-quota 5 minutes of GPU a day on a free account source write-gated verified email and an account older than 30 days source
Binding	volume-quota 5 minutes a day is 300 seconds, and a single diffusion call can eat a tenth of it
Interface	rest one https endpoint per Space serving a Gradio app; no SSH and no shell
Tooling	Gradio on the Space SDK; huggingface_hub
Composes with	5 minutes of GPU a day for a Gradio demo, which is an inference budget and not a training budget
Reliability	quota exhaustion drops the caller down the GPU queue, and the xlarge size costs double quota for the same wall-clock
Notes	The best VRAM for free and the smallest time budget. It is a serving allowance for a public demo, not a compute allowance for a job.
Evidence rung	documented the daily quotas, the account gate and the VRAM are quoted from the ZeroGPU docs today.
Checked	2026-10-25 (live); stale on 2026-11-24
Parsed numbers	`$.extras.m10h_usd` price USD0.0 / month `$.extras.m24h_usd` price USD0.0 / month `$.limits[0].value` limit:volume-quota 5.0 minutes

Evidence (8)

date	source	what it proved
2026-10-06	huggingface.co/docs/hub/en/spaces-zerogpu	free accounts in good standing may host up to 2 ZeroGPU Spaces with 5 minutes of GPU per day
2026-10-22	huggingface.co/docs/hub/en/spaces-overview	free personal accounts in good standing can host up to 2 Gradio Spaces running on ZeroGPU
2026-10-22	huggingface.co/docs/hub/spaces-zerogpu	'Free personal accounts: accounts in good standing (verified email, account older than 30 days) can host up to 2 ZeroGPU Spaces for free', with a usage table row 'Free account \| 5 minutes \| Medium', quoted
2026-10-24	huggingface.co/docs/hub/en/spaces-zerogpu	quoted: '\| Free account \| 5 minutes \| Medium \|' in the daily GPU quota table
2026-10-24	huggingface.co/docs/hub/en/spaces-zerogpu	quoted: 'Free personal accounts: accounts in good standing (verified email, account older than 30 days) can host up to 2 ZeroGPU Spaces for free.'
2026-10-24	huggingface.co/docs/hub/en/spaces-zerogpu	quoted: '\| `large` (default) \| Half NVIDIA RTX Pro 6000 Blackwell \| 48GB \| 1× \|'
2026-10-24	huggingface.co/docs/hub/en/spaces-zerogpu	quoted: 'Included daily quota resets exactly 24 hours after your first GPU usage.'
2026-10-24	huggingface.co/docs/hub/en/spaces-zerogpu	quoted, bold markers in the original: 'PRO, Team, and Enterprise users can continue using ZeroGPU Spaces beyond their included daily quota by consuming pre-paid credits at the rate of $1 per 10 minutes of GPU time.'

More fields (16)

{
 "shape": {
  "vcpu": null,
  "ram_gib": 16,
  "disk_gib": 50,
  "gpu_class": "NVIDIA RTX Pro 6000 Blackwell (half)",
  "gpu_count": 1,
  "note": "48 GB VRAM on the default 'large' size; the 16 GB / 50 GB figures are the CPU Basic envelope the Space runs in"
 },
 "billing_model": "free",
 "session_cap": {
  "wall_clock_min": null,
  "idle_reap_s": null,
  "restart_required": false,
  "note": "no session cap; 5 minutes of GPU a day"
 },
 "keep_rate": null,
 "persistent_storage": {
  "included_gib": null,
  "price_gib_month": "not persistent",
  "durability": "none",
  "survives_stop": false,
  "survives_destroy": false,
  "max_retention": "the Space lifetime"
 },
 "network_out": {
  "allow_outbound": true,
  "allow_inbound_listen": true,
  "ports": "one https endpoint per Space",
  "static_ip": false,
  "ipv6": false,
  "per_session_bytes": null
 },
 "free_tier_kind": "free-forever-with-quota",
 "census_rows": [
  "hf-spaces-zerogpu"
 ],
 "wave2": {
  "verdict_change": "trial-only -> qualifies-with-caveat on 2026-10-22"
 },
 "term": "none",
 "m10h_usd": 0.0,
 "m24h_usd": 0.0,
 "price_arithmetic": "USD 0.00 for 5 minutes a day; USD 1.00 per 10 minutes past the quota on a paid plan",
 "unattended": true,
 "unattended_mechanism": "the Space serves requests without a browser tab; each call gets a GPU for its duration",
 "units": "RAM and disk are quoted in the vendor's own decimal GB, not GiB. The ram_gib and disk_gib keys carry the vendor's number unconverted, so where a vendor writes 30 GB the key holds 30 and the true value is about 27.9 GiB. No conversion was performed."
}

Unknown or withheld values (1)

$.limits_note withheld

Corrections

2026-10-22 Hugging Face Spaces has a free lane, and the ZeroGPU allowance recurs daily

Hugging Face Storage Buckets via S3 gateway (s3.hf.co) — as an annex/DVC backend Storage

caveatinstitution?card?Rper-object-cap

huggingface.co/docs/hub/storage-buckets-s3 · academic · storage/hugging-face-storage-buckets-via-s3-gateway-as-an-annex · JSON · compare

Free tier	Public repos/buckets: 'Best-effort' — no published numeric ceiling on a Free account (private is a hard 100 GB). Storage plans table: Free user or org = Public 'Best-effort', Private 100 GB. HF is on the known list, but the S3 gateway is new and turns HF into a generic S3 endpoint, which is the part that matters for this slot. unit GB; metered on stored-bytes
Price	none stated
Per-object cap	single file will exceed 500GB (500000000000 bytes)
Retention	not stated
Egress	Free and CDN-backed: 'Egress and CDN included at no extra cost'. Served via CloudFront, so downloads are cacheable. This is the single best egress term of any LFS-adjacent host found.
Identity	institution (placeholder) inferred from the category 'academic'; the source record states no identity cost, so this is a class assumption, not evidence
Limits	per-object-cap single file will exceed 500GB source Bucket = not Git, so no git LFS semantics — you get S3 PUT/GET only. Hard limit: 'no single file will exceed 500GB' for Hub repos (buckets: gateway multipart, docs recommend 2 GB multipart chunks). Recommended <100k files and <10k entries per folder; >200 GB files discouraged due to CloudFront not caching them. Region pinned us-east-1. Gateway does not parse aws-chunked trailing checksums — must set request_checksum_calculation=when_required.
Binding	per-object-cap the per-object cap sets the shard size and therefore the object count; it bounds the total more tightly than any headline quota
Interface	s3, git, cli s3
Tooling	Any S3 client: AWS CLI, boto3, s5cmd; `hf` CLI (`hf sync`, `hf buckets`); python `huggingface_hub`. For this slot: `git annex initremote type=S3 host=s3.hf.co region=us-east-1 requeststyle=path` (annex needs path-style, which the gateway requires anyway) and `dvc remote add ... ; dvc remote modify myremote endpointurl https://s3.hf.co/<namespace>`. S3 creds are derived from an HF token via 'Generate S3 credentials' (HFAK... key).
Composes with	Bucket-per-purpose, unlimited buckets per namespace, unlimited free orgs. Because buckets are NON-versioned, git-annex supplies the versioning (hash-keyed objects) and DVC supplies the pointer layer — that combination gives you a versioned 500GB+ dataset store on a free, CDN-backed S3 endpoint. Dedupe is chunk-level on the Xet backend, so re-pushing near-identical files is cheap.
Reliability	Free public storage is explicitly 'best-effort' and HF says 'Beyond the first few gigabytes, please use this resource responsibly', with a mitigation the page describes as 'in place to prevent abuse', meaning a limit on excess use. That is the softest possible enforcement language, but also the least contractual. Gateway is single-region (us-east-1).
Notes	Public repos/buckets: 'Best-effort' — no published numeric ceiling on a Free account (private is a hard 100 GB). Storage plans table: Free user or org = Public 'Best-effort', Private 100 GB. HF is on the known list, but the S3 gateway is new and turns HF into a generic S3 endpoint, which is the part that matters for this slot.
Evidence rung	reported One page read moves it to documented.
Checked	2026-10-21 (inherited); stale on 2026-11-04
Parsed numbers	`$.limits[0].value` limit:per-object-cap 500.0 GB

Evidence (4)

date	source	what it proved
2026-10-21	huggingface.co/docs/hub/storage-buckets-s3	endpoint `https://s3.hf.co/<namespace>`, region us-east-1, path addressing, HFAK credentials from an HF token
2026-10-21	huggingface.co/docs/hub/storage-limits	Free user or org: Public 'Best-effort', Private 100GB; 'no single file will exceed 500GB'
2026-10-21	huggingface.co/storage	'Egress and CDN included at no extra cost'
2026-10-21	s3.hf.co/	probe: GET -> HTTP 403 with x-amz-request-id (live S3-compatible gateway answering)

Unknown or withheld values (5)

$.free_tier.notes process
$.identity_cost.kind identity-defaulted
$.per_object_cap.notes not-extracted
$.retention.notes not-extracted
$.retention.stated not-extracted

Hugging Face Inference Compute

paidemailno cardD

huggingface.co/docs/inference-providers/pricing · GPU cloud · compute/hugging-face-inference · JSON · compare

Free tier	The monthly credits table now reads: "Free Users \| None" and "PRO Users \| $2.00", with the note that free users can only use Inference Providers "by purchasing credits, or by upgrading to PRO". unit none; metered on wall-clock; none
Price	none stated
Per-object cap	not applicable
Retention	not applicable
Egress	not stated
Identity	email not published
Binding	none nothing binds because nothing is granted: the free lane on Inference Providers is now zero credits a month
Interface	unknown inferred from the tooling the provider documents; not probed
Tooling	not stated
Composes with	The USD 2 a month that PRO buys is the cheapest documented way to exceed the 5-minute ZeroGPU day, so a single PRO subscription is the paid upgrade path out of both free limits.
Reliability	no reliability event recorded
Notes	A refutation of the figure that circulated through 2025, where a free account received a small monthly inference credit. The free lane is now explicitly zero and the monthly credit starts at PRO.
Evidence rung	documented
Checked	2026-10-22 (live); stale on 2026-11-21

Evidence (2)

date	source	what it proved
2026-10-22	huggingface.co/docs/inference-providers/pricing	quoted monthly credits table: Free Users None, PRO Users USD 2.00, Team or Enterprise USD 2.00 per seat
2026-10-22	huggingface.co/docs/inference-providers/pricing	quoted: credits are general-purpose compute credits spendable on Inference Providers, Inference Endpoints, upgraded Spaces hardware including ZeroGPU beyond quota, and Jobs, credited every month

More fields (1)

{
 "wave2": {
  "class": "refutation-monthly-credit-removed",
  "free_tier_kind": "none",
  "shape": {
   "vcpu": null,
   "ram_gib": null,
   "gpu_class": null,
   "gpu_count": null
  }
 }
}

Unknown or withheld values (10)

$.egress.metered egress-unknown
$.egress.terms not-stated
$.identity_cost.detail process
$.interface interface-unknown
$.per_object_cap.notes process
$.per_object_cap.stated not-applicable
$.reliability_note none-observed
$.retention.notes process
$.retention.stated not-applicable
$.tooling not-stated

Hugging Face Jobs (hf-sandbox backend) Compute

paidemail?card?Drequest-rate

huggingface.co/docs/hub/en/jobs-pricing · platform host · compute/huggingface-jobs · JSON · compare

Free tier	no free compute allowance of its own: a Job bills by the minute while it is Starting or Running, at hardware rates from $0.01 an hour for CPU Basic to $23.50 an hour for eight L40S unit none; metered on none
Price	none stated
Per-object cap	not stated
Retention	not stated
Egress	not stated
Identity	email (placeholder) not published
Limits	request-rate billed by the minute only while a Job is Starting or Running, from $0.01 an hour for CPU Basic source per-object-cap 50 GB of ephemeral storage on CPU Basic, up to 6,500 GB on eight L40S source request-rate a monthly spending limit can be set per organisation or resource group source not stated
Binding	request-rate
Interface	unknown
Tooling	not stated
Composes with	not stated
Reliability	no reliability event recorded
Evidence rung	documented the Jobs pricing page was read on 2026-10-22 and publishes the hardware ladder and the minute billing. It states no free allowance for a Job, so the record's free-tier claim is not established for compute; whatever free tier the Hub has lives elsewhere
Checked	2026-10-23 (live); stale on 2026-11-22
Parsed numbers	`$.limits[1].value` limit:per-object-cap 50.0 GB

Evidence (2)

date	source	what it proved
2026-10-02	huggingface.co/docs/hub/en/jobs-pricing	not published
2026-10-22	huggingface.co/docs/hub/en/jobs-pricing	'Like hardware, it is billed by the minute, only while the Job is Starting or Running' with the hardware table: 'CPU Basic 2 vCPU 16 GB 50 GB $0.01' an hour, 'Nvidia T4 - small 4 vCPU 15 GB 16 GB 50 GB $0.40', up to '8x Nvidia L40S 192 vCPU 1534 GB 384 GB 6500 GB $23.50', quoted

More fields (14)

{
 "talaria_m10h_usd": 12.0,
 "talaria_m24h_usd": 16.2,
 "nemo_m10h_usd": null,
 "keep_rate": 0.0,
 "keep_rate_basis": "first-party: t a flat rate per Job, in addition to the hardware price: Product Hourly Price Exposed ports $0.01 Like hardwa",
 "billing_model": "unclassified",
 "shape": {
  "vcpu": 2,
  "ram": 16,
  "disk": 50,
  "gpu_class": null,
  "gpu_count": null,
  "note": "CPU Basic at $0.01 an hour; GPU shapes from a T4 at $0.40 to eight L40S at $23.50"
 },
 "session_cap": {
  "wall_clock_min": null,
  "idle_reap_s": null,
  "restart_required": false,
  "note": "a Job bills only while Starting or Running, and the page documents a timeout setting; the client bounds it"
 },
 "persistent_storage": null,
 "network_out": null,
 "price_note": null,
 "free_tier_kind": "none",
 "ps": {
  "gib": 50,
  "price": "ephemeral only",
  "durability": "ephemeral",
  "on_stop": false,
  "on_destroy": false,
  "retention": "the Job's lifetime"
 },
 "net": null
}

Unknown or withheld values (16)

$.chaining not-stated
$.egress.metered egress-unknown
$.egress.terms not-stated
$.evidence[0].proved process
$.extras.price_note process
$.identity_cost.detail defaulted
$.identity_cost.kind identity-defaulted
$.interface interface-unknown
$.interface_detail withheld
$.limits_note not-stated
$.per_object_cap.notes process
$.per_object_cap.stated not-stated
$.reliability_note none-observed
$.retention.notes not-extracted
$.retention.stated not-stated
$.tooling not-stated

Hugging Face Spaces (CPU Basic) Compute

paidcardcardDwrite-gateddisagrees

huggingface.co/docs/hub/en/spaces-overview · platform host · compute/huggingface-spaces-cpu-basic · JSON · compare

Sibling records disagree:

verdict: compute/huggingface-spaces says qualifies-with-caveat, compute/huggingface-spaces-cpu-basic says paid-only
identity_kind: compute/huggingface-spaces says email, compute/huggingface-spaces-cpu-basic says card

Free tier	CPU Basic is listed at 2 vCPU / 16 GB for FREE, but creating a Gradio or Docker Space requires a paid plan; a free account may host up to 2 Gradio Spaces on ZeroGPU instead unit hours; metered on wall-clock; The hardware is free on the price list; the right to create a compute Space is not.
Price	none stated
Per-object cap	not applicable: the unit is a session, not an object
Retention	50 GB of disk per Space, explicitly 'not persistent'; a Space sleeps after inactivity and rebuilds from the repo
Egress	no egress allowance published for Spaces
Identity	card a free account gets static Spaces; a Gradio or Docker Space needs PRO at USD 9 a month, so a card is the price of any compute Space. The ZeroGPU exception needs a verified email and an account older than 30 days.
Limits	write-gated PRO at USD 9 a month to create a Gradio or Docker Space on a personal account source per-object-cap 16 GB RAM, 2 CPU cores and 50 GB of not-persistent disk per Space source this record exists so the free hardware row is not mistaken for a free machine
Binding	write-gated the gate is a person paying USD 9, not a counter; the hardware would be free if the Space could be created
Interface	ssh the Space URL over HTTPS
Tooling	the Space SDK; git push to the Space repository
Composes with	none on a free account: the create step is gated
Reliability	a free-hardware Space sleeps after inactivity and wakes on request, so the first visitor after a quiet period waits for a cold start
Notes	This is the sharpest price-list trap: the hardware row says FREE and the same page's docs say a paid plan is required to use it. A reader who only reads the price table gets it wrong.
Evidence rung	documented the plan gate, the shape and the disk are quoted from the Spaces docs today, and the price list agrees that CPU Basic is FREE. The two pages together are the finding.
Checked	2026-10-25 (live); stale on 2026-11-24
Parsed numbers	`$.extras.m10h_usd` price USD9.0 / month `$.extras.m24h_usd` price USD9.0 / month `$.limits[1].value` limit:per-object-cap 16.0 GB

Evidence (7)

date	source	what it proved
2026-10-22	huggingface.co/docs/hub/en/spaces-overview	CPU Basic is listed as 2 vCPU / 16 GB at FREE, while creating a Gradio or Docker Space requires a paid plan
2026-10-22	huggingface.co/docs/hub/en/spaces-overview	Static Spaces are free for everyone; Gradio and Docker Spaces need PRO for personal accounts
2026-10-24	huggingface.co/docs/hub/en/spaces-overview	quoted: 'Static Spaces are free for everyone. Gradio and Docker Spaces run on compute and require a paid plan to create: PRO for personal accounts, Team or Enterprise for organizations.'
2026-10-24	huggingface.co/docs/hub/en/spaces-overview	quoted: 'Each Spaces environment is limited to 16GB RAM, 2 CPU cores and 50GB of (not persistent) disk space by default.'
2026-10-24	huggingface.co/docs/hub/en/spaces-overview	quoted: 'The default CPU Basic hardware has no hourly cost, but creating a Space that runs on compute (Gradio or Docker) requires a paid plan, while Static Spaces are free for everyone.'
2026-10-24	huggingface.co/pricing	quoted exactly, including the page's escaped hyphens: '\| CPU Basic \| 2 vCPU \| 16 GB \| \- \| \- \| FREE \|'
2026-10-24	huggingface.co/pricing	quoted: 'Upgrade your HF experience for $9/month.' — the PRO price that the compute-Space gate costs

More fields (15)

{
 "shape": {
  "vcpu": 2,
  "ram_gib": 16,
  "disk_gib": 50,
  "gpu_class": null,
  "gpu_count": null,
  "note": "vendor states 16GB and 50GB in decimal; kept as stated"
 },
 "billing_model": "free",
 "session_cap": {
  "wall_clock_min": null,
  "idle_reap_s": null,
  "restart_required": false,
  "note": "no session cap; sleeps on inactivity and wakes on request"
 },
 "keep_rate": null,
 "persistent_storage": null,
 "network_out": {
  "allow_outbound": true,
  "allow_inbound_listen": true,
  "ports": "one https endpoint per Space",
  "static_ip": false,
  "ipv6": false,
  "per_session_bytes": null
 },
 "free_tier_kind": "feature-gated",
 "census_rows": [
  "hf-spaces-cpu-basic",
  "huggingface-spaces"
 ],
 "term": "none",
 "m10h_usd": 9.0,
 "m24h_usd": 9.0,
 "price_arithmetic": "USD 9.00 a month for PRO, which is the flat cost of the right to create one compute Space; the hardware itself is USD 0.00",
 "unattended": true,
 "unattended_mechanism": "a Space runs as a server and stays up while it has traffic; it is a daemon, not a job, and sleeps when idle",
 "units": "RAM and disk are quoted in the vendor's own decimal GB, not GiB. The ram_gib and disk_gib keys carry the vendor's number unconverted, so where a vendor writes 30 GB the key holds 30 and the true value is about 27.9 GiB. No conversion was performed."
}

Hugging Face Spaces Dev Mode Compute

paidcardcardR

huggingface.co/docs/hub/spaces-dev-mode · platform host · compute/huggingface-spaces-dev-mode · JSON · compare

Free tier	Spaces Dev Mode is part of the PRO or Team and Enterprise plans unit none; metered on wall-clock; Dev Mode exposes an SSH endpoint, and it is paid
Price	none stated
Per-object cap	not applicable
Retention	the Space persists; Dev Mode is a session on it
Egress	not published
Identity	card PRO at USD 9 a month
Limits	write-gated Dev Mode is part of PRO or Team and Enterprise source it is not free
Binding
Interface	ssh SSH into a running Space while Dev Mode is on
Tooling	SSH, the Space SDK, a shell in the container
Composes with	the only SSH path into a Space, and it needs PRO
Reliability	no reliability event recorded
Notes	Recorded so the SSH claim is not repeated as free.
Evidence rung	reported re-read live on 2026-10-22 for the free-tier and limit figures
Checked	2026-10-23 (live); stale on 2026-11-06

Evidence (2)

date	source	what it proved
2026-10-02	huggingface.co/docs/hub/spaces-dev-mode	'The Spaces Dev Mode is part of PRO or Team & Enterprise plans.'
2026-10-22	huggingface.co/docs/hub/spaces-dev-mode	read live: Dev Mode requires PRO or higher

More fields (8)

{
 "shape": null,
 "billing_model": "wall-clock",
 "session_cap": null,
 "keep_rate": null,
 "persistent_storage": null,
 "network_out": {
  "allow_outbound": true,
  "allow_inbound_listen": true,
  "ports": "SSH while Dev Mode is on",
  "static_ip": false,
  "ipv6": null,
  "per_session_bytes": null
 },
 "free_tier_kind": "none",
 "census_rows": [
  "hf-dev-mode"
 ]
}

Unknown or withheld values (4)

$.egress.metered egress-unknown
$.per_object_cap.notes process
$.per_object_cap.stated not-applicable
$.reliability_note none-observed

Hugging Face Inference Providers Inference

paidcardcardU

router.huggingface.co/v1 · inference endpoint, platform allocation · inference/huggingface-inference · JSON · compare

Free tier	no monthly credits on a free account. 'Free users can use Inference Providers by purchasing credits, or by upgrading to PRO'; monthly credits are granted to PRO and to Team or Enterprise organizations unit none; metered on none; the pricing doc's eligibility row reads 'Free Users: yes (credits purchase required)'. So the router is reachable to a free account and only by paying, which is the same shape as a paid-only product with a free door
Price	none stated
Per-object cap	not applicable: an inference endpoint stores nothing for the caller
Retention	not applicable
Egress	the response is the payload; no separate egress term is published
Identity	card a Hugging Face account plus purchased credits, or a PRO subscription; a free account is refused by price rather than by key
Limits	request-rate a monthly inference credit allowance on the free account; the amount was not read the monthly credits granted to PRO, Team and Enterprise are shared among all members of an organisation and applied before any pay-as-you-go usage is billed; none of them is available to a free account
Binding	none no free allowance exists to bind: the free account can call the router only against purchased credits, so there is no ceiling on a free tier to name
Interface	rest
Tooling	any OpenAI-compatible client
Composes with	pairs with an agent through the router; the scarce link is the monthly credit, and the 401-as-HTML shape means a body-classifying client fails before it reads the quota
Reliability	the router returns its 401 as an HTML document rather than JSON, so a probe that classifies on body shape mis-handles it
Notes	It does not: the pricing documentation grants monthly credits to PRO and to Team or Enterprise only, and a free user reaches the router by purchasing credits. The verdict moves from qualifies-with-caveat to paid-only. The 401-as-HTML finding stands and is unchanged. Hugging Face also sells Jobs containers in the compute topic on the same account.
Evidence rung	unverified the eligibility text, the PRO and organisation credit grant and the router's 401 shape are all from pages read 2026-10-22. The 401 was re-probed at 2026-10-22T04:42:43Z and returned an HTML document rather than JSON.
Checked	2026-10-23 (live); stale on 2026-11-06

Evidence (2)

date	source	what it proved
2026-10-22	probe:https://router.huggingface.co/v1/chat/completions	an unauthenticated POST returned 401 with a 3087-byte body, an HTML document, not JSON. The endpoint exists and gates on a key; no key existed here, so no quota was read
2026-10-22	huggingface.co/docs/inference-providers/pricing	'PRO users and Team or Enterprise organizations receive monthly credits to use with Inference Providers. Free users can use Inference Providers by purchasing credits, or by upgrading to PRO'; the eligibility table reads 'Free Users — yes (credits purchase required)'

Unknown or withheld values (1)

$.retention.stated not-applicable

Hugging Face Spaces (disk check only) Storage

fails: quotaemail?card?R

huggingface.co/spaces · paas · storage/hugging-face-spaces · JSON · compare

Free tier	50 GB of disk per Space — but per HF's own docs the Space disk is NOT persistent. Persistent storage is the separate Storage Buckets product. unit GB; metered on stored-bytes
Price	none stated
Per-object cap	not stated
Retention	until the instance is recycled; not persistent
Egress	Free via the Hub CDN for public Spaces; no egress meter.
Identity	email (placeholder) not published
Binding	none no constraint of a named kind was recorded; the verdict rests on the quota sentence rather than on a named limit
Interface	rest, cli rest
Tooling	`huggingface_hub` / `hf` CLI; Spaces can mount buckets as volumes; Gradio/Docker Spaces are HTTP servers by construction.
Composes with	Not a storage play: each Space's 50 GB dies on restart. The real HF storage is buckets + public repos (see the HF dossier) — Spaces is only useful as the free server in front of them.
Reliability	Docs are explicit: 'This disk space is ephemeral, meaning its content will be lost if your Space restarts or is stopped.'
Notes	50 GB of disk per Space — but per HF's own docs the Space disk is NOT persistent. Persistent storage is the separate Storage Buckets product.
Evidence rung	reported One page read moves it to documented.
Checked	2026-10-21 (inherited); stale on 2026-11-04

Evidence (2)

date	source	what it proved
2026-10-21	huggingface.co/docs/hub/spaces-storage	'Every Space comes with a small amount of disk storage. This disk space is ephemeral'
2026-10-21	huggingface.co/docs/hub/spaces-overview	'50GB of (not persistent) disk space by default'; Gradio/Docker Spaces 'require a paid plan to create'

More fields (1)

{
 "dead_end": {
  "failure_class": "quota",
  "why": "50 GB of disk per Space — but per HF's own docs the Space disk is NOT persistent. Persistent storage is the separate Storage Buckets product.",
  "retry_when": null,
  "first_recorded": "2026-10-21"
 }
}

Unknown or withheld values (7)

$.extras.dead_end.retry_when process
$.free_tier.notes process
$.identity_cost.detail defaulted
$.identity_cost.kind identity-defaulted
$.per_object_cap.notes not-extracted
$.per_object_cap.stated not-extracted
$.retention.notes not-extracted

hugging face spaces disk (as persistent storage) Storage

fails: quotaemail?card?R

huggingface.co/docs/hub/spaces-storage · rejected-provider · storage/hugging-face-spaces-disk · JSON · compare

Free tier	50 GB sounds like a finding until you read the second sentence: 'This disk space is ephemeral'. Anything written to a Space is lost on restart, and creating a Gradio/Docker Space now needs a unit none; metered on none
Price	none stated
Per-object cap	not recorded
Retention	not recorded
Egress	not recorded
Identity	email (placeholder) not published
Binding	none no constraint of a named kind was recorded; the verdict rests on the quota sentence rather than on a named limit
Interface	unknown
Tooling	not recorded
Composes with	not recorded
Reliability	50 GB sounds like a finding until you read the second sentence: 'This disk space is ephemeral'. Anything written to a Space is lost on restart, and creating a Gradio/Docker Space now needs a
Notes	50 GB sounds like a finding until you read the second sentence: 'This disk space is ephemeral'. Anything written to a Space is lost on restart, and creating a Gradio/Docker Space now needs a
Evidence rung	reported
Checked	2026-10-21 (inherited); stale on 2026-11-04

Evidence (1)

date	source	what it proved
2026-10-21	huggingface.co/docs/hub/spaces-storage	50 GB sounds like a finding until you read the second sentence: 'This disk space is ephemeral'. Anything written to a Space is lost on restart, and creating a Gradio/Docker Space now needs a

More fields (1)

{
 "dead_end": {
  "failure_class": "retention",
  "why": "50 GB sounds like a finding until you read the second sentence: 'This disk space is ephemeral'. Anything written to a Space is lost on restart, and creating a Gradio/Docker Space now needs a",
  "retry_when": null,
  "first_recorded": "2026-10-21"
 }
}

Unknown or withheld values (15)

$.chaining not-recorded
$.confidence_note withheld
$.egress.terms not-recorded
$.extras.dead_end.retry_when process
$.free_tier.notes process
$.identity_cost.detail defaulted
$.identity_cost.kind identity-defaulted
$.interface interface-unknown
$.interface_detail withheld
$.limits_note process
$.per_object_cap.notes not-recorded
$.per_object_cap.stated not-recorded
$.retention.notes process
$.retention.stated not-recorded
$.tooling not-recorded

hugging face storage buckets (re-check only — not a discovery) Storage

fails: quotaemail?card?R

huggingface.co/docs/hub/storage-limits · rejected-provider · storage/hugging-face-storage-buckets · JSON · compare

Free tier	Recorded per the brief's instruction to re-verify HF numbers, not to claim it. The 2026 shape is materially different from the brief's entry: Free is 'Best-effort' public + 100 GB private; P unit none; metered on none
Price	none stated
Per-object cap	not recorded
Retention	not recorded
Egress	not recorded
Identity	email (placeholder) not published
Binding	none no constraint of a named kind was recorded; the verdict rests on the quota sentence rather than on a named limit
Interface	unknown
Tooling	not recorded
Composes with	not recorded
Reliability	Recorded per the brief's instruction to re-verify HF numbers, not to claim it. The 2026 shape is materially different from the brief's entry: Free is 'Best-effort' public + 100 GB private; P
Notes	Recorded per the brief's instruction to re-verify HF numbers, not to claim it. The 2026 shape is materially different from the brief's entry: Free is 'Best-effort' public + 100 GB private; P
Evidence rung	reported
Checked	2026-10-21 (inherited); stale on 2026-11-04

Evidence (3)

date	source	what it proved
2026-10-21	huggingface.co/docs/hub/storage-limits	Recorded per the brief's instruction to re-verify HF numbers, not to claim it. The 2026 shape is materially different from the brief's entry: Free is 'Best-effort' public + 100 GB private; P
2026-10-21	huggingface.co/docs/hub/storage-buckets-s3	cited as the basis of the rejection: Recorded per the brief's instruction to re-verify HF numbers, not to claim it. The 2026 shape is materially different from the brief's entry: Free is 'Best-effort' public + 100 GB private; P
2026-10-21	s3.hf.co	cited as the basis of the rejection: Recorded per the brief's instruction to re-verify HF numbers, not to claim it. The 2026 shape is materially different from the brief's entry: Free is 'Best-effort' public + 100 GB private; P

More fields (1)

{
 "dead_end": {
  "failure_class": "quota",
  "why": "Recorded per the brief's instruction to re-verify HF numbers, not to claim it. The 2026 shape is materially different from the brief's entry: Free is 'Best-effort' public + 100 GB private; P",
  "retry_when": null,
  "first_recorded": "2026-10-21"
 }
}

Unknown or withheld values (15)

$.chaining not-recorded
$.confidence_note withheld
$.egress.terms not-recorded
$.extras.dead_end.retry_when process
$.free_tier.notes process
$.identity_cost.detail defaulted
$.identity_cost.kind identity-defaulted
$.interface interface-unknown
$.interface_detail withheld
$.limits_note process
$.per_object_cap.notes not-recorded
$.per_object_cap.stated not-recorded
$.retention.notes process
$.retention.stated not-recorded
$.tooling not-recorded