# OpenRAIL

> Source: https://aiwiki.ai/wiki/openrail
> Updated: 2026-11-17
> Fact-checked: 2026-11-17
> Categories: AI Ethics, AI Policy & Regulation, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "OpenRAIL." aiwiki.ai, 17 Nov 2026. https://aiwiki.ai/wiki/openrail
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**OpenRAIL** (Open Responsible AI Licenses, also written Open RAIL) is a family of licenses for [machine learning](https://aiwiki.ai/wiki/machine_learning) models, source code, data and applications that gives royalty-free permission to use, modify and redistribute the licensed artifact, forbids a listed set of uses, and requires those use restrictions to be carried into the terms of every redistributed copy and derivative.[1][2] OpenRAIL licenses are the openly licensed subclass of the Responsible AI Licenses (RAIL), a family published since 2019 by the RAIL Initiative, a volunteer group of AI researchers and lawyers.[3][1][4]

The best-known members are model licenses, marked with the suffix "-M": the BigScience RAIL License v1.0 for the [BLOOM](https://aiwiki.ai/wiki/bloom) [large language model](https://aiwiki.ai/wiki/large_language_model) (May 2022), the CreativeML Open RAIL-M license for [Stable Diffusion](https://aiwiki.ai/wiki/stable_diffusion) v1 (August 2022), the CreativeML Open RAIL++-M license used for Stable Diffusion 2 and [SDXL](https://aiwiki.ai/wiki/sdxl), and the BigCode OpenRAIL-M license for [StarCoder](https://aiwiki.ai/wiki/starcoder) (May 2023).[5][6][7][8][9] A study of the [Hugging Face](https://aiwiki.ai/wiki/hugging_face) model hub found that in January 2024 RAIL licenses were attached to 41,700 model repositories, 24.0% of those that declared a license, second only to the Apache family.[10]

Because they restrict fields of use, OpenRAIL licenses are not open source licenses under the Open Source Definition, as the initiative's own FAQ states.[1][11] Lawyers who work on open licensing have questioned whether the restrictions can be enforced, and in April 2026 the Free Software Foundation added the RAIL licenses to its list of nonfree licenses.[12][13] Since late 2023 several developers have chosen other terms for new models: [Stability AI](https://aiwiki.ai/wiki/stability_ai) moved first to non-commercial licences and then to its own Community License, and others have released under [Apache 2.0](https://aiwiki.ai/wiki/apache_license) or [MIT](https://aiwiki.ai/wiki/mit_license).[14][15][16]

## Origins: the RAIL Initiative

The first Responsible AI Licenses appeared in early 2019. VentureBeat reported that February that RAIL had been formed by researchers from [Google](https://aiwiki.ai/wiki/google), [Microsoft](https://aiwiki.ai/wiki/microsoft) and IBM who met through the ACM Future of Computing Academy. It identified Danish Contractor of IBM Research in New Delhi as a co-founder, named Julia Haines of Google and Brent Hecht of Northwestern University, a co-chair of the academy, among those involved, and said that patent lawyer Christopher Hines had written the two initial agreements over about a year of discussion with the group.[4] Daniel McDuff, then at [Microsoft Research](https://aiwiki.ai/wiki/microsoft_research), is identified as a member of the initiative in its 2022 paper and sits on its steering committee with Contractor, Haines, Hecht, Joseph Lindley, Hanlin Li, Carlos Muñoz Ferrandis, Nick Vincent, Tim Korjakow, Jenny Lee and Jesse Josua Benjamin.[17][18] A later paper by the group dates the proposal of licenses with behavioral-use clauses to 2018.[10]

The two 2019 documents were an end-user license, meant to sit beside a vendor's existing terms, and a source code license.[19][20] The source code license grants a royalty-free copyright license and then lists uses the software and its derivatives may not be put to, grouped under four headings: surveillance, computer generated media, health care and "criminal" uses such as predicting the likelihood that a person will commit a crime.[21] The initiative wrote that it expected the idea to be contested: "We expect that this will be controversial".[19]

RAIL describes itself as a volunteer-driven coordinating group that works on standardizing the licenses, on tools for choosing and tracking them, and on developer education.[3]

| Date | Event |
| --- | --- |
| January 2019 | First prototype RAIL licenses released (end-user and source code)[22][4] |
| January 2020 | IEEE P2840 working group on responsible AI licensing established[22] |
| 2020 | AAAI Spring Symposium on responsible AI licensing[2] |
| May 2022 | BigScience RAIL License v1.0, co-developed with RAIL, written for BLOOM[5][2] |
| June 2022 | "Behavioral Use Licensing for Responsible AI" presented at ACM FAccT[17] |
| August 18, 2022 | Naming convention and the "Open" prefix introduced[2] |
| March 2023 | AI Pubs RAIL licenses published for authors of research papers[23] |
| March 19, 2024 | RAIL License Generator announced[24] |
| November 12, 2024 | IEEE 2840-2024, Standard for Responsible AI Licensing, approved (published May 2, 2025)[25] |
| April 22, 2026 | Free Software Foundation lists RAIL as nonfree[13] |

The IEEE standard came out of the working group formed in 2020, which Contractor chairs. It does not define what a license is. It requires conforming licenses to contain clauses on the use of specific technical artifacts.[25]

## Naming scheme

In August 2022 Contractor, Muñoz Ferrandis, Lee and McDuff published a naming convention. Under it, a license is a RAIL license if it includes behavioral-use restrictions and requires downstream use and redistribution to include, at minimum, the same restrictions.[2] Letters after the name identify which artifact the restrictions cover, and are combined in the fixed order D-A-M-S.

| Suffix | Artifact covered by the use restrictions |
| --- | --- |
| D | Data: the datasets used to pretrain or train a model |
| A | Application or service: executable software, including API access |
| M | Model: learned weights and parameters, including checkpoints |
| S | Source: the source code and scripts of the AI system |

A license restricting only the model is a RAIL-M license, one restricting model and source is RAIL-MS, and one covering all four would be RAIL-DAMS.[2]

The prefix "Open" marks the licenses in which, in the convention's words, "the licensor offers the licensed artifact at no charge and allows licensees to re-license such artifact or any subsequent derivative works as they choose, as long as the Use Restrictions similarly apply to the relicensed artifacts and its subsequent derivatives".[2] A RAIL license that charges for the artifact or forbids downstream licensing does not take the prefix, so a proprietary or research-only license can be RAIL without being OpenRAIL.[2][1] The license generator that RAIL launched in 2024 offers three types on this basis: OpenRAIL, Research RAIL (research use only) and plain RAIL.[24]

## Licenses in the family

| License | Dated | Steward | Notable use |
| --- | --- | --- | --- |
| Responsible AI End-User License (a RAIL-A license) | 2019 | RAIL Initiative | Sample license[20] |
| Responsible AI Source Code License (an Open RAIL-S license) | 2019 | RAIL Initiative | Sample license[20] |
| BigScience RAIL License v1.0 | May 19, 2022 | BigScience | [BLOOM](https://aiwiki.ai/wiki/bloom)[5] |
| BigScience Open RAIL-M | August 18, 2022 | BigScience | General-purpose model license[26][27] |
| CreativeML Open RAIL-M | August 22, 2022 | CompVis and Stability AI | [Stable Diffusion](https://aiwiki.ai/wiki/stable_diffusion) v1.x[6][1] |
| OpenRAIL++-M template | Not separately dated (keeps the CreativeML date of August 22, 2022) | Yannic Kilcher | Blank template[28][1] |
| CreativeML Open RAIL++-M | November 24, 2022; July 26, 2023 | Stability AI | Stable Diffusion 2.x; [SDXL](https://aiwiki.ai/wiki/sdxl) 1.0[7][8] |
| AI Pubs OpenRAIL-M and OpenRAIL-S; AI Pubs Research-Use RAIL-M and RAIL-S | March 2023 | RAIL Initiative | Models and code released with papers[23] |
| BigCode OpenRAIL-M v1 | May 2023 | BigCode | [StarCoder](https://aiwiki.ai/wiki/starcoder), StarCoder2[29][9][30] |
| SIL AI RAIL-M | Not stated | SIL International | NLP models for indigenous languages[1] |

The model licenses descend from one text. The BigScience RAIL License v1.0 was drafted for the BLOOM models within the BigScience workshop, by authors from its Legal and Ethical and Model Governance working groups, and the RAIL FAQ calls it the first OpenRAIL-M license and the first open license written specifically for a machine learning model.[31][1] BigScience then generalized it into the BigScience Open RAIL-M so that it could be attached to any model, including multimodal generative ones.[26] The CreativeML Open RAIL-M, whose copyright notice names [Robin Rombach](https://aiwiki.ai/wiki/robin_rombach) and Patrick Esser, adapted that text for [text-to-image models](https://aiwiki.ai/wiki/text-to-image_models): according to the RAIL FAQ it has a new preamble and drops two of the BigScience restrictions.[6][1] Comparing the texts, the two that are missing are the requirement to disclose that content is machine generated and the ban on impersonating others.[27][6] RAIL's writing also treats research-only model licenses with behavioral clauses, such as [Meta](https://aiwiki.ai/wiki/meta_ai)'s licenses for OPT-175B and SEER, as RAIL licenses that are not open.[26]

The "++" variant removes a sentence that drew complaints. Clause 7 of the BigScience and CreativeML licenses says that "You shall undertake reasonable efforts to use the latest version of the Model."[6] The RAIL FAQ explains that the clause was conceived for critical model failures, but that some in the community read it as an obligation to keep upgrading, even to a worse model. Yannic Kilcher published a template without it, and Stability AI's CreativeML Open RAIL++-M for Stable Diffusion 2 and SDXL likewise keeps only the licensor's right to restrict usage that violates the license.[1][28][7][8]

BigCode, a research collaboration led by Hugging Face and [ServiceNow](https://aiwiki.ai/wiki/servicenow), released StarCoder in May 2023 under what its announcement called "an improved version of the OpenRAIL license" that "simplifies the process for companies to integrate the model into their products".[9] The BigCode OpenRAIL-M v1 adds a restriction on generating malware, narrows the medical-advice clause, and lets recipients share the model "under any license of your choice that does not contradict the restrictions in Attachment A of this License Agreement".[29]

Hugging Face lists six identifiers for the family in its license metadata: `openrail`, `bigscience-openrail-m`, `creativeml-openrail-m`, `bigscience-bloom-rail-1.0`, `bigcode-openrail-m` and `openrail++`.[32] The SPDX License List, by contrast, has no entry for any RAIL license as of November 2026.[33]

## What the licenses say

The CreativeML Open RAIL-M shows the structure the model licenses share.[6]

### Grant

The intellectual property sections are modeled on the Apache License 2.0.[34][1] Each contributor grants "a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model", together with a patent license that terminates if the licensee files patent litigation over the model.[6] The definitions are specific to machine learning. "Derivatives of the Model" covers any model created by transferring the weights, parameters, activations or output of the original, "including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model". A model produced by [knowledge distillation](https://aiwiki.ai/wiki/knowledge_distillation) is therefore a derivative. The training data is not licensed, and "Distribution" includes offering the model as a hosted service or API.[6]

### Use restrictions

Paragraph 5 makes the list in Attachment A binding: "You cannot use the Model and the Derivatives of the Model for the specified restricted uses."[6] The CreativeML attachment has eleven items. Among them, the licensee agrees not to use the model:

- "In any way that violates any applicable national, federal, state, local or international law or regulation"
- "For the purpose of exploiting, harming or attempting to exploit or harm minors in any way"
- "To generate or disseminate verifiably false information and/or content with the purpose of harming others"
- "To defame, disparage or otherwise harass others"
- "To provide medical advice and medical results interpretation"[6]

The remaining items cover personal information that can be used to harm an individual, fully automated decisions that affect legal rights, discrimination based on social behavior, personal characteristics or legally protected categories, exploitation of vulnerable groups, and generating information for use in the administration of justice, law enforcement, immigration or asylum processes.[6] The BigScience licenses list thirteen restrictions and the BigCode license fourteen.[5][29] The 2019 source code license had been organized differently, by application area: surveillance, synthetic media, health care and crime prediction.[21]

### Passing restrictions downstream

A recipient may redistribute the model or a derivative under different license terms, commercial ones included, on one condition that defines the family. Paragraph 4 states: "Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model".[6] The preamble puts it this way: derivative versions "will always have to include - at minimum - the same use-based restrictions as the ones in the original license". Paragraph 5 adds that "You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph".[6] RAIL compares the mechanism to the copyleft and share-alike clauses of open source and Creative Commons licenses.[1] What propagates is the list of forbidden uses: the license does not oblige anyone to publish a derivative or to keep the same license for it, and the flow-down provision does not apply to the accompanying source code.[6]

### Outputs and remote restriction

On generated content, the license says that "Licensor claims no rights in the Output You generate using the Model" and that "No use of the output can contravene any provision as stated in the License."[6] Clause 7 reserves to the licensor "the right to restrict (remotely or otherwise) usage of the Model in violation of this License", a provision kept in the "++" and BigCode versions.[6][8][29]

## Research behind the licenses

The framework is set out in "Behavioral Use Licensing for Responsible AI" by Danish Contractor, Daniel McDuff, Julia Katherine Haines, Jenny Lee, Christopher Hines, Brent Hecht, Nicholas Vincent and Hanlin Li. It was first posted to [arXiv](https://aiwiki.ai/wiki/arxiv) in November 2020 and published at the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT '22), held in Seoul on June 21-24, 2022.[17] The paper starts from the observation that ethical guidelines for AI "can help set norms and shape policy" but "are not easily enforceable", and argues that license terms can make behavioral conditions on code, models and data legally enforceable. Two case studies turn recommendations from published AI ethics guidelines and from AI FactSheets into license clauses, and the paper ends with an exemplar license. It also discusses limits, including the cost of enforcement for volunteer developers, and cites the JSON license's "The Software shall be used for Good, not Evil" to illustrate how hard enforceable restrictions are to write.[17]

A second paper, by McDuff, Contractor and ten co-authors, was posted to arXiv in February 2024 as "On the Standardization of Behavioral Use Clauses and Their Adoption for Responsible Licensing of AI" and appeared at [ICML](https://aiwiki.ai/wiki/icml) 2024 under the title "Position: Standardization of Behavioral Use Clauses is Necessary for the Adoption of Responsible Licensing of AI".[10][35] It combines interviews with four practitioners who had released projects under such licenses, a coding of 77 clauses from seven licenses into 22 clause types, and a count of licenses on the Hugging Face model hub. The authors found that almost all the licenses restricted discrimination, disinformation and violations of law, and argued for "standardized customization": a common core with domain-specific additions, supported by tooling.[10] The RAIL License Generator, announced in March 2024, implements that idea, stamping each generated license with a creation date, a unique identifier and the version of the generator that produced it.[24]

A May 2025 preprint by McDuff, Korjakow, Kevin Klyman and Contractor reported that more than 300 customized licenses had been created with the generator in 14 months, and argued that the next need was tooling to track adherence to behavioral-use clauses.[36]

## Legal and community status

### Not open source

Clause 6 of the Open Source Definition, maintained by the [Open Source Initiative](https://aiwiki.ai/wiki/open_source_initiative) (OSI), reads: "The license must not restrict anyone from making use of the program in a specific field of endeavor."[11] A list of prohibited uses conflicts with it directly, and the drafters say so themselves. BigScience's introduction to its license answered the question in its FAQ: "This is not an open source license according to the Open Source Initiative definition, because it has some restrictions on the use of the model."[31] The RAIL FAQ repeats the point in capitals, while maintaining that the licenses are open in the sense of allowing reuse, distribution, commercialization and adaptation outside the restricted uses. It also warns that an OpenRAIL-licensed model cannot be redistributed under an open source license, because the use restrictions must stay attached.[1]

The OSI's [Open Source AI Definition](https://aiwiki.ai/wiki/open_source_ai_definition) keeps the same principle for AI systems: its first freedom is to "Use the system for any purpose and without having to ask for permission."[37] In its 2023 Deep Dive: AI webinar series, the OSI hosted a session titled "Should OpenRAIL licenses be considered OS AI Licenses?" with McDuff, Contractor, Lee and the open source lawyer Luis Villa.[38] Krzysztof Siewicz, writing for the Free Software Foundation's licensing and compliance team in April 2026, announced that the foundation had added the RAIL licenses to its list of nonfree licenses, on the ground that "any use restriction in a software license makes the program nonfree".[13]

### Enforceability

Three lawyers who work on software licensing published detailed critiques in 2022 and 2023. Each questioned how the restrictions would be enforced.

Luis Villa, writing for Tidelift in November 2022, noted that it is unclear whether copyright applies to a trained model at all, in which case a defendant's first argument would be that no license is needed. He also pointed out that a well-resourced party could retrain a model from scratch to escape the license, that each customized appendix is in effect a different license with uncertain compatibility, and that prohibitions on conduct such as harassment are unlikely to affect people who do not read license agreements, whereas restrictions aimed at regulated organizations (medical advice, for example) might.[34]

Kyle Mitchell examined the flow-down clause in January 2023 and found that it obliges a redistributor to make the restrictions enforceable, not to enforce them, and does not say who may sue a downstream violator: "Enforceable by who? Not clear."[39]

Kate Downing argued in July 2023 that the licenses are styled as copyright licenses and so depend on copyright in the weights: "there is deep skepticism in the IP world that the models are even copyrightable. If not, the license is fully unenforceable." She added that the restrictions are vague enough that a court might decline to enforce some of them, that conditioning a copyright license on conduct outside the scope of copyright risks a copyright misuse defense in the United States, and that a licensor has no contract with users who received the model from an intermediary. Her conclusion was that licensing cannot substitute for regulation, and that a developer who believes a model could be seriously misused should not publish it.[12]

The initiative has acknowledged these limits. Its 2022 paper lists enforcement as an open problem, and its FAQ presents litigation as the exception: a licensor who finds a violation is advised first to contact the licensee, and then to ask the hosting platform to assess the case and take the offending model down.[17][1] It also argues that the restrictions act as a deterrent and as a statement of norms even before any enforcement.[40][24]

### Criticism and support

The Free Software Foundation's objection is one of principle. Siewicz wrote that use-restricting licenses hand private licensors an authority that belongs to democratic lawmaking, that "Most, if not all, activities prohibited in RAIL are already illegal in most countries", and that the licenses require none of the things users need to control machine learning systems, such as training inputs and source code. The post said the foundation acted because it still received questions about the licenses "despite these licenses having not gained much popularity".[13]

Supporters start from the view that models differ from code. Carlos Muñoz Ferrandis, then at Hugging Face, wrote in August 2022 that recent events had sent the message that "Openness is not enough", and that OpenRAIL licenses "could be for open and responsible ML what current open software licenses are to code and Creative Commons to general content: a widespread community licensing tool".[40] Hugging Face supported the approach and added the licenses to its license metadata.[40][32]

Some reviewers took a middle position. Villa concluded that the family does not meet the Open Source Definition but, because Stable Diffusion and BLOOM made it hard to avoid, might be "the most important license of the next 3-5 years", and urged the open source legal community to work with its drafters.[34] GitLab's Robin Schulman wrote in 2023: "We see RAIL-M as an exciting but flawed development in the AI model licensing space."[41] Stephen O'Grady of RedMonk wrote in 2026 that "while OpenRAIL licenses cannot be considered Permissive in the OSI-approved sense, in spirit they embody many of the same values".[42]

Assessments of individual models reflect the same split. In the validation exercise for its [Open Source AI Definition](https://aiwiki.ai/wiki/open_source_ai_definition), the OSI said that BLOOM and StarCoder2, both released under OpenRAIL licenses, "would probably pass if they changed their licenses/legal terms".[52] The paper that introduced the [Model Openness Framework](https://aiwiki.ai/wiki/model_openness_framework) lists OpenRAIL among the licenses it does not treat as open source because of their usage restrictions, while the openness assessments by Andreas Liesenfeld and Mark Dingemanse of Radboud University, discussed at [openwashing](https://aiwiki.ai/wiki/openwashing), accept a RAIL license for the top grade of their license dimension.[53][54]

## Adoption

Stable Diffusion drove most of the early uptake: McDuff and colleagues found text-to-image to be the most common task among RAIL-licensed repositories.[10] Measured shares depend on what is counted, and the published figures are not directly comparable.

| Date | Finding | Source |
| --- | --- | --- |
| September 2022 | 209 RAIL-licensed model repositories on Hugging Face, 1% of licensed repositories | McDuff et al.[10] |
| September 2022 to January 2023 | Open RAIL share of licensed repositories rose from 0.5% to 7%; permissive open source licenses above 80% | Keller and Bonato, Open Future[43] |
| January 2023 | 4,035 repositories, 10% | McDuff et al.[10] |
| January 2024 | 41,700 repositories, 24.0% of 174,163 licensed; Apache 38.1%, MIT 17.5%, Creative Commons 7.5% | McDuff et al.[10] |
| 2025 (preprint of May 2025) | Licenses with behavioral-use clauses on 12.1% of about 650,000 licensed repositories; open source licenses 61.5% | McDuff et al. preprint[36] |
| May 2026 | OpenRAIL the third most common license family after Apache and MIT in a scan of about 2.9 million models | O'Grady, RedMonk[42] |

The 2024 paper added that, counted by downloads instead of repositories, RAIL licenses ranked fourth, behind Apache, MIT and Creative Commons, and that "some of the most popular models are not using RAIL licenses".[10] The 2025 figure uses a wider definition: its authors make no distinction between RAIL-branded licenses and other licenses with behavioral-use clauses, such as those of [Llama](https://aiwiki.ai/wiki/llama), [Gemma](https://aiwiki.ai/wiki/gemma) and DeepSeek.[36] Keller and Bonato concluded in 2023 that Open RAIL licenses had so far had limited effect in a field dominated by permissive licenses.[43] The RedMonk scan found that almost seventy percent of Hugging Face models carried no license at all, and called OpenRAIL "the largest single non-OSI, non-model-specific license category in this dataset".[42]

The Hugging Face model listing gave the following counts per license tag on November 17, 2026, out of roughly 3.24 million models.[44]

| License tag | Models |
| --- | --- |
| `openrail` | 41,853 |
| `creativeml-openrail-m` | 33,464 |
| `openrail++` | 17,745 |
| `bigscience-bloom-rail-1.0` | 3,008 |
| `bigscience-openrail-m` | 2,879 |
| `bigcode-openrail-m` | 2,543 |
| All six OpenRAIL tags | 101,492 |
| `apache-2.0` (for comparison) | 596,387 |
| `mit` (for comparison) | 215,916 |

The two sets of figures are not measured the same way (the 2024 study grouped licences into families, the 2026 counts are individual Hub tags), but they point in one direction. In January 2024 the study counted 41,700 RAIL repositories against 66,402 in the Apache family, a ratio of roughly 0.63. In November 2026 the six OpenRAIL tags together (101,492 models) came to roughly 0.17 of the `apache-2.0` tag alone (596,387). The family grew in absolute numbers and shrank relative to Apache.[10][44]

### Developers moving to other terms

Stability AI, whose Stable Diffusion releases made the family widely known, stopped using it for new image models. SDXL 1.0 was released under the CreativeML Open RAIL++-M in July 2023.[8] SDXL Turbo, published in November 2023, already carried a Stability non-commercial license instead of an OpenRAIL license.[51] Stable Diffusion 3 Medium followed on June 12, 2024, under the Stability Non-Commercial Research Community License.[45] On July 5, 2024, saying that SD3's original commercial licensing had "caused some confusion and concern in the community", the company introduced the Stability AI Community License. It allows research, non-commercial and commercial use free of charge for individuals and organizations with annual revenue up to USD 1 million, and requires a paid enterprise license above that.[14] [Stable Diffusion 3.5](https://aiwiki.ai/wiki/stable_diffusion_3_5) was released under the Community License in October 2024.[46] The Community License gates commercial use by revenue and refers to an acceptable use policy, a different design from OpenRAIL's fixed list with no commercial threshold.[14]

[Black Forest Labs](https://aiwiki.ai/wiki/black_forest_labs), launched in August 2024 by a team that describes itself as the creators of [latent diffusion](https://aiwiki.ai/wiki/latent_diffusion) and the Stable Diffusion models, released [FLUX.1](https://aiwiki.ai/wiki/flux_1) [schnell] under Apache 2.0 and FLUX.1 [dev] under a non-commercial license.[15]

Among language model developers, [DeepSeek](https://aiwiki.ai/wiki/deepseek) made a similar change. Its DeepSeek License Agreement, version 1.0 of October 23, 2023, used for models such as [DeepSeek Coder](https://aiwiki.ai/wiki/deepseek_coder), reproduces the OpenRAIL-M structure, including the preamble's statement that derivatives must carry "at minimum" the same use-based restrictions and an Attachment A of prohibited uses.[47] [DeepSeek-R1](https://aiwiki.ai/wiki/deepseek_r1), released on January 20, 2025, carries the MIT license.[16]

Other adopters stayed. BigCode released StarCoder2 in February 2024 under the same BigCode OpenRAIL-M license as its predecessor.[30]

## Comparison with other use-restricted licenses

Other licenses also pair downloadable weights with conditions on use, most of them written by a single vendor for its own models. [Open-weight license comparison](https://aiwiki.ai/wiki/open_weight_license_comparison) covers their commercial terms in detail.

| License | Use restrictions | Other conditions |
| --- | --- | --- |
| OpenRAIL-M licenses | Fixed list in an attachment to the license; must be passed to derivatives[6] | No revenue or user threshold |
| Llama 4 Community License ([Meta](https://aiwiki.ai/wiki/meta_ai)) | Acceptable Use Policy incorporated by reference[48] | Separate license needed above 700 million monthly active users; "Built with Llama" notice[48] |
| Gemma Terms of Use ([Google](https://aiwiki.ai/wiki/google)) | Prohibited Use Policy incorporated by reference; must be passed on as an enforceable provision[49] | Google reserves the right to restrict usage it reasonably believes violates the terms[49] |
| Stability AI Community License | Acceptable use policy[14] | Free up to USD 1 million annual revenue[14] |

The Gemma terms use nearly the same flow-down wording as the OpenRAIL-M licenses, requiring distributors to "include the use restrictions referenced in Section 3.2 as an enforceable provision in any agreement" governing Gemma or its derivatives.[49] One difference is where the restrictions sit: an OpenRAIL attachment is part of a dated license text, whereas the Llama and Gemma agreements point to a separate policy document.[6][48][49] The 2024 RAIL study counted the [Llama 2](https://aiwiki.ai/wiki/llama_2) license as a behavioral-use license, on 3,566 Hugging Face repositories in January 2024.[10]

Google has since changed course for new releases: [Gemma 4](https://aiwiki.ai/wiki/gemma_4), announced on April 2, 2026, was the first Gemma generation published under Apache 2.0.[50]

## See also

- [Open-weight license comparison](https://aiwiki.ai/wiki/open_weight_license_comparison)
- [Open weights](https://aiwiki.ai/wiki/open_weights)
- [Open-source AI](https://aiwiki.ai/wiki/open_source_ai)
- [Apache License 2.0](https://aiwiki.ai/wiki/apache_license)
- [MIT License](https://aiwiki.ai/wiki/mit_license)
- [OpenMDW](https://aiwiki.ai/wiki/openmdw)
- [Model card](https://aiwiki.ai/wiki/model_card)
- [Responsible AI](https://aiwiki.ai/wiki/responsible_ai)

## References

1. Responsible AI Licenses (RAIL). "FAQ". licenses.ai. https://www.licenses.ai/faq-2
2. Contractor, Danish; Muñoz Ferrandis, Carlos; Lee, Jenny; McDuff, Daniel. "From RAIL to Open RAIL: Topologies of RAIL Licenses". licenses.ai, August 18, 2022. https://www.licenses.ai/blog/2022/8/18/naming-convention-of-responsible-ai-licenses
3. Responsible AI Licenses (RAIL). "Responsible AI Licenses". licenses.ai. https://www.licenses.ai/
4. Johnson, Khari. "RAIL debuts license agreements for the responsible use of AI". VentureBeat, February 11, 2019. https://venturebeat.com/2019/02/11/rail-debuts-license-agreements-for-the-responsible-use-of-ai/
5. BigScience. "BigScience RAIL License v1.0". Hugging Face, May 19, 2022. https://huggingface.co/spaces/bigscience/license
6. Rombach, Robin; Esser, Patrick; and contributors. "CreativeML Open RAIL-M". GitHub (CompVis/stable-diffusion), August 22, 2022. https://github.com/CompVis/stable-diffusion/blob/main/LICENSE
7. Stability AI and contributors. "CreativeML Open RAIL++-M License". GitHub (Stability-AI/stablediffusion, archived copy), November 24, 2022. https://web.archive.org/web/20240710140638/https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/LICENSE-MODEL
8. Stability AI. "CreativeML Open RAIL++-M License" (SDXL 1.0). Hugging Face, July 26, 2023. https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md
9. Hugging Face. "StarCoder: A State-of-the-Art LLM for Code". Hugging Face blog, May 4, 2023. https://huggingface.co/blog/starcoder
10. McDuff, Daniel; Korjakow, Tim; Cambo, Scott; Benjamin, Jesse Josua; Lee, Jenny; Jernite, Yacine; Muñoz Ferrandis, Carlos; Gokaslan, Aaron; Tarkowski, Alek; Lindley, Joseph; Cooper, A. Feder; Contractor, Danish. "On the Standardization of Behavioral Use Clauses and Their Adoption for Responsible Licensing of AI". arXiv:2402.05979, February 7, 2024. https://arxiv.org/abs/2402.05979
11. Open Source Initiative. "The Open Source Definition". opensource.org. https://opensource.org/osd
12. Downing, Kate. "AI Licensing Can't Balance 'Open' with 'Responsible'". Law Offices of Kate Downing, July 13, 2023. https://katedowninglaw.com/2023/07/13/ai-licensing-cant-balance-open-with-responsible/
13. Siewicz, Krzysztof. "RAIL: Nonfree and unethical". Free Software Foundation, April 22, 2026. https://www.fsf.org/blogs/licensing/rail-are-nonfree-and-unethical
14. Stability AI. "Community License". stability.ai, July 5, 2024. https://stability.ai/news-updates/license-update
15. Black Forest Labs. "Announcing Black Forest Labs". bfl.ai, August 1, 2024. https://bfl.ai/announcing-black-forest-labs/
16. DeepSeek. "DeepSeek-R1" (model card, MIT license). Hugging Face, January 20, 2025. https://huggingface.co/deepseek-ai/DeepSeek-R1
17. Contractor, Danish; McDuff, Daniel; Haines, Julia Katherine; Lee, Jenny; Hines, Christopher; Hecht, Brent; Vincent, Nicholas; Li, Hanlin. "Behavioral Use Licensing for Responsible AI". 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT '22), pp. 778-788, June 2022. doi:10.1145/3531146.3533143. https://arxiv.org/abs/2011.03116
18. Responsible AI Licenses (RAIL). "Team". licenses.ai. https://www.licenses.ai/team
19. Responsible AI Licenses (RAIL). "About". licenses.ai. https://www.licenses.ai/about
20. Responsible AI Licenses (RAIL). "AI Licenses". licenses.ai. https://www.licenses.ai/ai-licenses
21. Responsible AI Licenses (RAIL). "Source Code License" (Responsible Artificial Intelligence Source Code License). licenses.ai. https://www.licenses.ai/source-code-license
22. Responsible AI Licenses (RAIL). "News". licenses.ai. https://www.licenses.ai/news
23. Responsible AI Licenses (RAIL). "AI Pubs RAIL Licenses". licenses.ai, March 2023. https://www.licenses.ai/blog/2023/3/3/ai-pubs-rail-licenses
24. Responsible AI Licenses (RAIL). "Announcing the RAIL License Generator". licenses.ai, March 19, 2024. https://www.licenses.ai/blog/2024/3/19/announcing-the-rail-license-generator
25. IEEE Standards Association. "IEEE 2840-2024: IEEE Standard for Responsible AI Licensing". standards.ieee.org. https://standards.ieee.org/ieee/2840/7673/
26. Muñoz Ferrandis, Carlos; Contractor, Danish. "BigScience Open RAIL-M License". licenses.ai, August 26, 2022. https://www.licenses.ai/blog/2022/8/26/bigscience-open-rail-m-license
27. BigScience. "BigScience Open RAIL-M License". August 18, 2022. https://static1.squarespace.com/static/5c2a6d5c45776e85d1482a7e/t/6308bb4bba3a2a045b72a4b0/1661516619868/BigScience+Open+RAIL-M+License.pdf
28. Kilcher, Yannic. "OpenRAIL++-M License Template". ykilcher.com. https://www.ykilcher.com/license
29. BigCode. "BigCode Open RAIL-M v1 License Agreement". Hugging Face. https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement
30. BigCode. "StarCoder2-15B" (model card, license: bigcode-openrail-m). Hugging Face, February 2024. https://huggingface.co/bigcode/starcoder2-15b
31. Contractor, Danish; Muñoz Ferrandis, Carlos. "The BigScience RAIL License". BigScience blog (archived copy), 2022. https://web.archive.org/web/20240128035240/https://bigscience.huggingface.co/blog/the-bigscience-rail-license
32. Hugging Face. "Licenses". Hugging Face Hub documentation. https://huggingface.co/docs/hub/repositories-licenses
33. SPDX. "SPDX License List". spdx.org. Accessed November 17, 2026. https://spdx.org/licenses/
34. Villa, Luis. "Evaluating the RAIL license family". Tidelift blog (now hosted by Sonar), November 1, 2022. https://www.sonarsource.com/blog/evaluating-the-rail-license-family/
35. McDuff, Daniel; et al. "Position: Standardization of Behavioral Use Clauses is Necessary for the Adoption of Responsible Licensing of AI". Proceedings of the 41st International Conference on Machine Learning, PMLR 235:35255-35266, 2024. https://proceedings.mlr.press/v235/mcduff24a.html
36. McDuff, Daniel; Korjakow, Tim; Klyman, Kevin; Contractor, Danish. "New Tools are Needed for Tracking Adherence to AI Model Behavioral Use Clauses". arXiv:2505.22287, May 28, 2025. https://arxiv.org/abs/2505.22287
37. Open Source Initiative. "The Open Source AI Definition 1.0". opensource.org. https://opensource.org/ai/open-source-ai-definition
38. Open Source Initiative. "Should OpenRAIL licenses be considered OS AI Licenses?". Deep Dive: AI webinar series. https://opensource.org/ai/webinars/should-openrail-licenses-be-considered-os-ai-licenses
39. Mitchell, Kyle E. "Open RAIL-M Is More Unclear Than Scary". /dev/lawyer, January 26, 2023. https://writing.kemitchell.com/2023/01/26/Open-RAIL-M-Unclear
40. Muñoz Ferrandis, Carlos. "OpenRAIL: Towards open and responsible AI licensing frameworks". Hugging Face blog, August 31, 2022. https://huggingface.co/blog/open_rail
41. Schulman, Robin. "RAIL-M is an imperfectly good start for AI model licenses". GitLab blog, July 25, 2023. https://about.gitlab.com/blog/rail-m-is-an-imperfectly-good-start-for-ai-model-licenses/
42. O'Grady, Stephen. "License Distribution on Hugging Face". RedMonk tecosystems, May 12, 2026. https://redmonk.com/sogrady/2026/05/12/hugging-face-licensing/
43. Keller, Paul; Bonato, Nicolò. "The Growth of responsible AI licensing". Open Future, February 7, 2023. https://openfuture.eu/publication/the-growth-of-responsible-ai-licensing/
44. Hugging Face. "Models" (listing filtered by license tag; counts read on November 17, 2026). https://huggingface.co/models?license=license:creativeml-openrail-m
45. Stability AI. "Stable Diffusion 3 Medium". stability.ai, June 12, 2024. https://stability.ai/news-updates/stable-diffusion-3-medium
46. Stability AI. "Introducing Stable Diffusion 3.5". stability.ai, October 22, 2024. https://stability.ai/news-updates/introducing-stable-diffusion-3-5
47. DeepSeek. "DeepSeek License Agreement, Version 1.0". GitHub (deepseek-ai/DeepSeek-Coder), October 23, 2023. https://github.com/deepseek-ai/DeepSeek-Coder/blob/main/LICENSE-MODEL
48. Meta. "Llama 4 Community License Agreement". llama.com. https://www.llama.com/llama4/license/
49. Google. "Gemma Terms of Use". Google AI for Developers, last modified April 1, 2026. https://ai.google.dev/gemma/terms
50. Castelly, Nia; casari, amanda; Lacombe, Olivier. "Gemma 4: Expanding the Gemmaverse with Apache 2.0". Google Open Source Blog, April 2, 2026. https://opensource.googleblog.com/2026/03/gemma-4-expanding-the-gemmaverse-with-apache-20.html
51. Stability AI. "SDXL-Turbo" (model card). Hugging Face. https://huggingface.co/stabilityai/sdxl-turbo
52. Open Source Initiative. "Open Source AI Definition: Frequently Asked Questions". opensource.org. https://opensource.org/ai/faq
53. White, Matt; Haddad, Ibrahim; Osborne, Cailean; Liu, Xiao-Yang Yanglet; Abdelmonsef, Ahmed; Varghese, Sachin; Le Hors, Arnaud. "The Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency, and Usability in Artificial Intelligence". arXiv:2403.13784. https://arxiv.org/abs/2403.13784
54. Liesenfeld, Andreas; Dingemanse, Mark. "Rethinking open source generative AI: open-washing and the EU AI Act". FAccT '24, June 2024. https://doi.org/10.1145/3630106.3659005

