Hugging Face model and dataset licensing on the Hub is a metadata and access-control layer. The Hub lets a repository owner declare a licence in the card, which is the repository’s README.md, so that other users know the permissions attached to the code or data, and it tells users to seek out and respect a project’s licence before using it.[1] The Hub does not grant or enforce those permissions itself. For asset managers the practical consequence is that every downloaded model and dataset carries its own licence, which may be permissive, copyleft, a use-restricted AI licence or a bespoke community licence, and which has to be read where the Hub shows it.
Editions
Licence identifiers
The Hub documents a list of identifiers that can be set in the card metadata. Software-style entries include Apache license 2.0 (apache-2.0), MIT (mit), the BSD family, the GNU GPL, LGPL and AGPL entries, MPL 2.0, EPL 1.0 and 2.0, EUPL 1.1 and 1.2 and the Unlicense.[1] Content and data entries include the Creative Commons family (cc-by-4.0, cc-by-sa-4.0, cc-by-nc-4.0, cc-by-nd-4.0 and the other variants), the Community Data License Agreements (cdla-sharing-1.0, cdla-permissive-1.0 and cdla-permissive-2.0), the Open Database License family, Open Data Commons entries and the Computational Use of Data Agreement.[1]
Model-specific entries are the part most relevant to AI governance. They include the OpenRAIL license family (openrail), BigScience OpenRAIL-M (bigscience-openrail-m), CreativeML OpenRAIL-M (creativeml-openrail-m), BigScience BLOOM RAIL 1.0, BigCode Open RAIL-M v1 and Open Rail++-M (openrail++).[1] They also include the Llama community licence identifiers (llama2, llama3, llama3.1, llama3.2, llama3.3 and llama4), the Gemma Terms of Use (gemma), Grok 2 Community License Agreement (grok2-community), the Apple Model License for Research, the Intel Research Use License Agreement, the FAIR Noncommercial Research License, the DeepFloyd IF Research License Agreement, the H Research License and the Open Model, Data & Weights License Agreement (openmdw-1.0 and openmdw-1.1).[1] The list also has unknown and other values.[1] See The licence list covers OpenRAIL, Llama, Gemma and Creative Commons identifiers.
Custom licences
When no identifier fits, the author sets license: other and adds the licence text to a LICENSE file in the repository, or contacts Hugging Face to add the licence to the list, and sets a name in license_name.[1] The model card documentation adds a license_link field, so a custom licence can point to a URL, with an example that names a bespoke model licence, and, where no URL exists, a LICENSE stored in the model repository.[2] See Licence other requires the licence text in the repository.
Display and discovery
The licence chosen in the model card metadata is displayed on the model page and users can filter models by licence.[2] For datasets, if the chosen licence uses a keyword from the licence table, the licence is displayed on the dataset page, and card metadata lets users filter and discover datasets.[3] See The licence shown on a model page can be used to filter models and A licence is declared in repository card metadata.
Metrics
None of the licence features is a billing metric. The identifiers in the metadata are labels on repositories, and access requests are per user and not metered. The catalog therefore records them as rules and programs. Seat and storage metrics that affect a licence manager appear in Hugging Face Hub plans, seats and storage.
Counting / floors
OpenRAIL and use-based restrictions
Hugging Face’s August 2022 article describes Open and Responsible AI licenses (OpenRAIL) as AI-specific licences that allow open access, use and distribution of AI artifacts while requiring responsible use.[7] It gives two defining features: the licences allow royalty-free access and flexible downstream use and redistribution, including derivatives, and they embed restrictions on use in identified critical scenarios.[7] The article also states that OpenRAIL licences require downstream adoption of the use-based restrictions in redistribution and in derivatives.[7] Compared with an OSI-style software licence, the restrictions are on how the model is used, not only on how the files are copied, so a compliance review has to check intended use against the licence. See OpenRAIL licences combine open use with use-based restrictions.
Llama-style and other community licences
The identifier list contains separate entries for the Llama 2, 3, 3.1, 3.2, 3.3 and 4 community licence agreements, Gemma and Grok 2.[1] Hugging Face’s documentation records the identifier but does not summarize the terms, so the terms of use, any use thresholds, acceptable-use conditions and attribution duties must be read from the licence text in the repository or from its publisher. Practically, a licence version identifier such as llama3.1 versus llama3.2 is a different licence and should be recorded as such.
Gated repositories
Authors can enable access requests for models and datasets. A repository with access requests is called gated, and users must agree to share their contact information (username and email address) with the authors to access the files.[4][5] Access requests are always granted to individual users rather than to entire organizations, so a company cannot approve itself once for all staff.[4] See Gated access is granted to individual users, not organizations.
By default access is granted automatically once the user has submitted the information, which the documentation calls automatic approval. With manual approval the author reviews pending requests and can accept or reject each one. A rejected user cannot access the model and cannot request access again, and an accepted user can be rejected at any time or moved back to pending.[4] A reset status revokes the previous decision and asks the user to agree to the gating terms again.[4] See Gated models use automatic or manual approval and Gated datasets work like gated models.
Authors can customize the form with extra_gated_fields in the model card metadata, with field types such as text, checkbox, date picker, country and select, and a prompt message with extra_gated_prompt. The documentation’s example asks for company and country and has a checkbox for non-commercial use only.[4] The information is available to the author in the repository settings and can be downloaded as an access report with the user, full name, status, email and request times.[4] Gating can be automated through an API, and the documentation notes use cases such as screening requests outside the Hub or conditioning access on a payment flow that happens outside the Hub.[4] See Gate forms can collect extra fields and a licence acknowledgement.
The Content Policy defines a gated repository as one whose repository and community content are visible to everyone, while access to the machine learning artifacts (data, model weights) requires accepting conditions in a click-through form or approval by the maintainers.[9] Team and Enterprise organizations can also create Gating Group Collections, in which one access request covers all models and datasets in a collection, provided the collection owner is a subscribed organization and all items belong to the same organization.[6] See Gating Group Collections cover several repositories with one request.
Terms of Service overlay
The Terms of Service add three rules that bear on licences. A user who sets a repository public grants each user a perpetual, irrevocable, worldwide, royalty-free, non-exclusive licence to use, display, publish, reproduce, distribute and make derivative works of the content through the services and functionalities.[8] Content that contains notice of a reasonable and customary licence, such as an open source licence, is intended to remain under that licence when further accessed, distributed or used.[8] Content downloaded from another user is used at the user’s own risk and subject to the Terms and the terms accompanying the content.[8] See A public repository grants each user a licence to the content, An open source licence notice on content continues to apply and Downloaded content is used at the user’s risk and subject to its own terms.
Virtualization & partitioning
Not applicable. A licence or gate applies to the repository, not to processors or virtual machines. How a model’s licence treats fine-tuned derivatives, hosted inference or redistribution depends on the licence text.
Cloud / BYOL
Running a gated or restricted-licence model on Hugging Face compute, such as a dedicated Inference Endpoint, does not replace the model’s licence. The compute is billed under the subscription and usage terms in Hugging Face compute billing, and the Terms of Service say that open source and similar licence terms are not replaced or overridden by the Terms.[8]
Programs
Gated access with automatic or manual approval is the only licence-related program on the Hub, together with Gating Group Collections for Team and Enterprise organizations.[4][6] The Content Policy lists gating as a tool that repository owners can use to review and approve or reject access to their artifacts.[9]
Out of scope
This article does not summarize the Llama, Gemma, Grok, OpenRAIL or other licences themselves, and it does not assess whether an organization’s use is permitted under any of them. It does not cover the licences of the Hugging Face open-source libraries, export-control rules for specific models or the training-data provenance of hosted models. Inventorying which licence identifier each downloaded model carries, recording the gate acceptance, and tracking licence version changes are practitioner steps and are not Hugging Face requirements.