Hugging Face

A platform for hosting and collaborating on open machine-learning models, datasets, and apps, with pay-as-you-go inference and compute.Reference profile

Hugging Face runs a platform for hosting open ML models, datasets and apps (Spaces) and maintains open-source libraries such as transformers, diffusers and huggingface_hub. The platform is closed and hosted; transformers is Apache-2.0. Vendor homepage figures: 2M+ models, 1M+ apps, 400,000+ organizations. Paid items (2026-10-26): PRO $9/month; Team and Enterprise from $20/user/month; per-TB storage; hourly Spaces hardware and Inference Endpoints; Inference Providers at provider rates. Latest transformers release as of 2026-10-31: v5.22.1 (October 30).

EditorialUpdated Oct 26, 202610 min read

What it is

Hugging Face's core product is the Hub, a hosting platform organized around Git repositories of three types: models, datasets and Spaces. Storage Buckets, a non-versioned storage type, joined them in 2026. Model repos hold weights and model cards, dataset repos hold training and evaluation data, and Spaces are demo apps that run in the browser, built with Gradio, Docker or static HTML. Public repos are hosted for free; private repos, extra compute and organization features are paid.

Beyond the Hub, the company maintains open-source libraries: transformers for model definitions and loading, diffusers for diffusion models, huggingface_hub as the Python client for the Hub, plus tokenizers, datasets, PEFT, TRL, Accelerate, smolagents and Transformers.js. transformers has about 168,000 GitHub stars and is Apache-2.0.

The paid business has three lines. Subscriptions: PRO for individuals, Team and Enterprise for organizations. Compute: Spaces hardware upgrades, Inference Endpoints for dedicated serving billed by the hour, and Jobs. Inference Providers: one API that routes requests to more than a dozen inference companies, among them Baseten, Cerebras, Fireworks, Groq, Together and Replicate, with Hugging Face passing their prices through.

On scale, the homepage claims 2M+ models, 1M+ apps and 500k+ datasets (as reported by the vendor). The August 2026 State of Open Models report gives finer numbers: between January and August, public model repos grew from 2.43 million to 2.96 million, datasets from 711,000 to 1 million and Spaces from 1.00 million to 1.44 million.

The contracting entity is Hugging Face, Inc. The terms of service require individual users to be at least 13 and organization accounts to be duly registered legal entities. Sign-up, repository hosting and community discussions all run on one account system, and the homepage's organization list shows each organization's plan: Google, Microsoft, Amazon and AI at Meta are labeled Enterprise, Grammarly and Writer Team.

Hugging Face charges you the same rates as the provider, with no additional fees. We just pass through the provider costs directly.
Inference Providers pricing docs

Core capabilities

Model distribution is the Hub's base. Replicate, in the Inference Providers launch post, calls the Hub the de facto home of open-source model weights (a partner's view). Model pages show the model card, files, download counts and license, and link straight to Inference Providers for a trial. The homepage's trending list mixes models from DeepSeek, Black Forest Labs, Lightricks and others, so the Hub is not tied to any one model maker.

Storage moved to Xet. From January 2025 the Hub migrated from Git LFS to Xet, a chunk-deduplicating backend: editing a large file uploads and downloads only the changed chunks rather than the whole file. The migration post says 500,000 repos holding 20 PB moved to Xet within six months (as reported by the vendor).

Storage Buckets, launched on March 10, 2026, are mutable object storage for checkpoints, optimizer states, logs and other intermediate files that change often and need no version history. Buckets expose an S3-compatible API, so AWS CLI, boto3 and similar tools work with them, and deleted files free space at once, unlike Git repos that keep every past version.

ZeroGPU lends GPUs to Spaces on demand. It runs on NVIDIA RTX Pro 6000 Blackwell cards, allocated when a function is called and released when it finishes. Free users can run ZeroGPU apps, and PRO users get eight times the free daily quota.

The formats and libraries reach beyond the Hub. Safetensors, the standard format for distributing weights on the Hub, moved to the PyTorch Foundation on April 8, 2026, with its trademark, repository and governance held by the Linux Foundation. Since September 2026 transformers can load llama.cpp GGUF quantized models directly.

Agent support is the 2026 focus. The hf CLI was rebuilt in June for coding agents, and the vendor's benchmark says a no-CLI baseline uses up to six times as many tokens on complex multi-step tasks (as reported by the vendor). In September the Hub began grouping reinforcement-learning environments under a tag and filter, and from October 20 transformers, diffusers and Kernels run natively on TPUs through TorchTPU. The late-October changelog adds passkey sign-in, 3D asset previews and external collaborators for private repos.

Who it's for / Not for

A fit for: researchers and developers who publish or look for open models and datasets. Public repos are free and unlimited, model cards and licenses sit on each page, and download and like counts are public.

A fit for: individuals building model demos or small apps. CPU Basic hardware is free; free accounts with a confirmed email that are at least 30 days old can host up to two ZeroGPU Spaces, and PRO accounts up to ten.

A fit for: organizations that manage model assets centrally. Team costs $20 per user per month and adds SAML and OIDC single sign-on, audit logs, resource-group permissions and a choice of storage region; Enterprise adds SCIM user sync and annual contracts.

A fit for: teams trying several inference providers from one account. Inference Providers passes provider prices through; PRO accounts include $2 of monthly credits, and Team and Enterprise include $2 per seat.

Not a fit for: steady, long-running production inference on a zero budget. ZeroGPU is shared, with daily quotas and queues, supports only the Gradio SDK, and is less compatible than Spaces with a dedicated GPU.

Not a fit for: teams with no capacity to vet content and licenses. Most models and datasets on the Hub come from third parties under varied licenses and with uneven quality; the mid-year report says about 85.6% of models have fewer than 200 lifetime downloads. The terms of service restrict use in some regions and for some purposes; see the terms for details.

How it compares

Against code-hosting platforms, the Hub also builds on Git repositories but is designed for large files: the Xet backend deduplicates in 64KB chunks, and model pages preview weight tensors, render model cards, show download statistics and call inference directly.

Against cloud object storage, public repos on the Hub are free and paid storage is priced per TB: from $12 per TB per month for public and $18 for private, falling to $8 and $12 above 500 TB, with egress and CDN included, as of 2026-10-26. The Hugging Face pricing page lists AWS S3 at $23 per TB and Backblaze Overdrive at $15 as reference points (as of 2026-10-26, as reported by the vendor).

Against inference aggregators, Inference Providers adds no markup: the pricing docs state that Hugging Face charges provider rates with no additional fees (as reported by the vendor). A custom provider key is also supported, in which case the provider bills directly. Providers include Baseten, Cerebras, Cohere, DeepInfra, Fireworks, Groq, Novita, Replicate and Together, and the catalog centres on open-weight models.

Against local inference tools, the GGML team behind llama.cpp joined in February 2026 while the project stays open source and community-led, and transformers now loads GGUF files, linking the two ecosystems.

Getting started

Browsing and downloading public models and datasets needs no account. A free account adds uploads, likes, discussions and access tokens for the API and the command line.

Programmatic access rests on two tools: the huggingface_hub Python library and the hf CLI built on it. The CLI covers downloads and uploads, repos and branches, Jobs, Buckets and Inference Endpoints. huggingface_hub v1.0, released in October 2025, brought breaking changes, and older projects need the migration guide when upgrading.

Models can be called three ways. Inference Providers suits pay-per-call trials billed through the Hugging Face account; Inference Endpoints suits dedicated deployments billed per instance-hour; Spaces suits shareable demos, starting on free CPU, with paid hardware or ZeroGPU for GPU work.

Organizations on the Team plan get single sign-on, resource groups and organization billing in one place. Organization credits for Inference Providers apply only when a request names the organization to be billed.

Pricing & license

Personal accounts are free. PRO costs $9 a month and adds 10x private storage, $2 of monthly inference credits and 8x ZeroGPU quota. For organizations, Team is $20 per user per month with 1 TB of private storage per seat. Enterprise is listed at $50 per user per month on the pricing page but as "custom pricing" on the Enterprise page; both statements stand as of 2026-10-26.

Storage is priced per TB per month: $12 public and $18 private at base, $10 and $16 above 50 TB, $9 and $14 above 200 TB, and $8 and $12 above 500 TB. Buckets are billed on stored volume too, with Enterprise plans billed on the deduplicated footprint.

Spaces hardware starts with free CPU Basic; CPU Upgrade is $0.03 an hour, one T4 $0.40, one A100 $2.50 and eight RTX PRO 6000 cards $22. Inference Endpoints CPU instances start at $0.03 an hour, one H100 costs $4.50 an hour and one B200 $9.25.

Inference Providers passes provider prices through. Included credits apply only to requests routed through Hugging Face; with a custom provider key, the provider bills directly.

On licensing, the Hub platform and paid services are closed hosted services; transformers is Apache-2.0. Every model and dataset on the Hub carries its own license, as marked on its repo.

Risks & limits

Availability. The terms of service restrict use in some regions and for some purposes; see the terms for details. Individual users must be at least 13. Changes to the terms and the privacy policy take effect ten days after posting, and continued use counts as acceptance.

Content risk. Models, datasets and Spaces on the Hub are mostly uploaded by third parties. The content policy, effective April 10, 2025, sets out banned content and the handling process, but repos are not reviewed one by one in advance. Licenses, data provenance and safety are left to users, and the homepage trending list has included apps labelled "uncensored".

Inconsistent figures. The homepage says Inference Providers gives access to 45,000+ models, while the pricing docs say 200+ models (both as reported by the vendor); Enterprise pricing differs between the pricing page and the Enterprise page.

Compute limits. ZeroGPU is shared, with daily quotas and queues; the xlarge size uses quota at twice the rate, and only the Gradio SDK is supported. Free accounts must be at least 30 days old to host ZeroGPU apps.

Data location. Only Team and Enterprise plans can pick the storage region for repo data; free and PRO accounts cannot. The privacy policy says personal information may be stored and processed in several countries.

Long tail. The mid-year report says about 85.6% of models have fewer than 200 lifetime downloads and 1.5% of repos account for 99.2% of downloads, so popularity and usability are not the same thing.

How it got here

The company open-sourced a PyTorch port of BERT in 2018, and transformers grew into the main model-definition library. On December 21, 2021, Hugging Face announced the acquisition of Gradio, one of the frameworks Spaces supports. On May 9, 2022, it raised a $100 million Series C led by Lux Capital.

On January 28, 2025, Inference Providers launched with fal, Replicate, SambaNova and Together AI. On April 14 the company announced the acquisition of open-source robot maker Pollen Robotics, describing it as its fifth acquisition after deals including Gradio and XetHub. A July 15 post reported Xet migration progress, with Xet the standard backend for new users and organizations since May. huggingface_hub reached v1.0 on October 27, and on December 1 the first transformers v5 release candidate arrived, with the post citing more than 3 million pip downloads a day (as reported by the vendor).

On February 20, 2026, the GGML and llama.cpp team joined Hugging Face; the announcement says the team keeps working full time on llama.cpp with autonomy over technical direction. Storage Buckets followed on March 10, Safetensors moved to the PyTorch Foundation on April 8, the hf CLI was rebuilt for agents on June 4, the State of Open Models mid-year report came out on August 14, transformers gained GGUF support on September 22, and the Hub began grouping RL environments on September 28.

In October, transformers v5.20.0 shipped on October 16 and v5.20.2 on October 20; TorchTPU support also landed on the 20th, passkey sign-in on the 21st, and external collaborators for private repos on the 23rd. The transformers repository has about 438 contributors with GitHub accounts, as of 2026-10-26.

Our verdict

Inference: a fit for teams that treat open models and datasets as everyday raw material. The Hub is the most concentrated venue for open-model releases, and Xet storage, Buckets and the hf CLI make uploading, syncing and versioning large files smoother than on general code platforms; public hosting is free, and paid storage is priced per TB with no egress charges.

Inference: a fit for teams comparing several inference providers from one account. Inference Providers passes prices through, credits work across several compute services, and moving from trials to dedicated Inference Endpoints stays on the same platform.

Inference: not a fit for teams that need strict control over model provenance, licensing and content safety but lack reviewers, since most Hub content comes from third parties. Also not a fit for running production on free ZeroGPU: quotas, queues and framework limits affect stability.

Inference: a fit for organizations that need single sign-on, audit logs and a choice of data location, provided per-seat Team pricing is acceptable; Enterprise pricing is stated inconsistently, so the contract quote is what counts.

Why it matters

Hugging Face is one of the main hubs for open models and datasets, and the home of open-source libraries such as transformers.

Sources

Media & third party

Updates & reading

NewsReutersOct 28, 2026

US Justice Department makes inquiries into Nvidia's $12.93 billion deal to buy Hugging Face

Reuters reported on Oct 28, citing two people familiar with the matter, that the US Justice Department has asked how Nvidia's $12.93 billion deal for open-source AI platform Hugging Face would affect competition, including whether it would give Nvidia a key distribution point for AI. CNBC later confirmed it. Nvidia declined to comment; the deal is expected to close in the first half of 2027.

2 sources

NewsXOct 26, 2026