Citation and evidence

Model Openness Framework

25 min full readUpdated 38 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI Policy & RegulationModel EvaluationOpen Source AI

Cite this article

The Model Openness Framework (MOF) is a classification system from the Linux Foundation's LF AI & Data Foundation that grades a released machine learning model on two things: how many of the artifacts from its development are published (completeness) and whether each of them carries an open licence (openness).[1][2] It lists 17 components of a model release, grouped as code, data and documentation, and sorts qualifying releases into three cumulative classes: Class III (Open Model), Class II (Open Tooling) and Class I (Open Science).[2] The framework was written in the foundation's Generative AI Commons working group, first published as an arXiv paper in March 2024, and issued as a stable specification, version 1.0, dated December 17, 2024.[1][2]

A companion web application, the Model Openness Tool (MOT) at isitopen.ai, takes a description of what a producer has released and under which licences, returns the class the release qualifies for, and keeps a public registry with badges.[3][4] Classification is self-reported, and no central body certifies it.[5] The authors present the framework as an answer to openwashing, the practice of calling a restrictively licensed or partly released model "open source".[1][6] The MOF is separate from the Open Source Initiative's Open Source AI Definition (OSAID), which borrowed the MOF's list of components for its own checklist but sets a single bar that a system either clears or does not.[7][8] In 2025, work on applying the MOF led to OpenMDW, a permissive licence intended to cover every component of a model release in one document.[9]

History

The framework was set out in "The Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency, and Usability in Artificial Intelligence", posted to arXiv on March 20, 2024. Its seven authors are Matt White and Ibrahim Haddad of the Linux Foundation, Cailean Osborne (then at the University of Oxford), Xiao-Yang Yanglet Liu of Columbia University, Ahmed Abdelmonsef and Sachin Varghese of the Generative AI Commons, and Arnaud Le Hors of IBM.[1][5] The paper was revised five times, and the sixth version is dated October 18, 2024.[1] Its acknowledgements thank reviewers from EleutherAI, the Open Source Initiative (Stefano Maffulli and Justin Colannino), Intel, HP, Futurewei and Cloud Geometry.[5]

LF AI & Data announced the framework on April 17, 2024, describing the three classes and promising an assessment tool and a badge programme.[10] A beta of the Model Openness Tool followed at the Linux Foundation's AI_dev Europe conference in Paris on June 19-20, 2024, where Haddad, the executive director of LF AI & Data, announced the launch in his keynote and White demonstrated the tool.[11][12] The Stack reported at the time that the tool had not yet identified a single model that met the top class.[13]

The Generative AI Commons then condensed the paper into "The Model Openness Framework (MOF) Specification", version 1.0, dated December 17, 2024 and edited by Le Hors. It restates the paper's requirements in RFC 2119 terms (MUST, SHOULD, MAY), is released under CC-BY-4.0, and describes itself as largely aligned with the sixth version of the paper. It lists four deviations: the configuration file is recommended instead of required, the data card may be folded into the model card, a per-component LICENSE file may cover components of mixed types, and the configuration file may use any machine-readable format, not only JSON.[2] A Japanese translation of the specification was published on April 7, 2025.[14] A version of the paper that also covers the OpenMDW licence, by White, Osborne, Liu, Keyi Wang and Varghese, was presented as a poster at a NeurIPS 2025 workshop on regulatable machine learning.[15]

No later version of the specification had been published as of November 2026. The tool's documentation still links version 1.0, and registry entries added in May 2026 declare framework version 1.0.[4][16]

DateEvent
2024-03-20First version of the paper posted to arXiv[1]
2024-04-17LF AI & Data announces the framework[10]
2024-06Beta of the Model Openness Tool launched at AI_dev Europe, Paris[11][12]
2024-10-18Sixth and latest version of the paper[1]
2024-10OSI finalizes the Open Source AI Definition 1.0, whose checklist uses the MOF component list[6][8]
2024-12-17MOF Specification 1.0[2]
2025-04-07Japanese translation of the specification[14]
2025-05-22Linux Foundation presents the OpenMDW licence as an outgrowth of MOF work[9]
2026-01Tool gains submission of entries by pull request from the evaluation form[16]
2026-05Le Hors presents the tool at Open Source Summit North America[6][17]
2026-05-28OpenMDW-1.1 released[18]

Expanded article table

Completeness and openness

The MOF keeps its two criteria apart. Completeness is about which artifacts are in the distribution. Openness is about the licence on each one.[6] The specification defines an open licence as "a license allowing unrestricted usage, study, modification, and redistribution for any purpose", and asks for a licence suited to the type of artifact: an open source licence for code, an open data licence for data, and an open content licence for documentation.[2]

Openness is binary for each component. The paper applies the open source movement's rule that there is no in-between: a component under a licence that is not open does not count toward any class. A producer may leave a component out and qualify for a lower class, provided everything that remains is openly licensed.[5] The paper names the Llama 2 licence, OpenRAIL and the AI2 ImpACT licences as examples that are not open source licences because they restrict use.[5] A model whose terms are too restrictive receives no class at all.[6]

Model parameters are treated as data. The specification says software licences such as Apache 2.0 and MIT "are not well suited" to weights and recommends an open data licence such as CDLA-Permissive-2.0, although its licence table still accepts OSI-approved licences for parameters.[2] The tool follows the same logic: it accepts an open licence of the wrong type but shows a warning, as it does for IBM Granite weights released under Apache 2.0.[4][19]

Components

The specification defines 17 components. The table follows its numbering, which encodes the class in which a component first appears (III.1 is the first Class III component, I.6 the last Class I component).[2]

No.ComponentTypeRequired fromLicence in specification 1.0
III.1Model architectureCodeClass IIIOSI-approved licence
III.2Final model parameters (checkpoints and optimizer state)DataClass IIIPreferred CDLA-Permissive-2.0; OSI-approved and open data licences accepted
III.3Technical reportDocumentationClass III (a research paper may stand in)Preferred CC-BY-4.0; open content licences accepted
III.4Evaluation resultsDocumentationClass IIIPreferred CC-BY-4.0; open content licences accepted
III.5Model cardDocumentationClass IIIPreferred CC-BY-4.0; open content licences accepted
III.6Data cardDocumentationClass IIIPreferred CC-BY-4.0; open content licences accepted
III.7Sample model outputsData or codeOptionalUnlicensed (shared without restrictions)
II.2Training, validation and testing codeCodeClass IIOSI-approved licence
II.3Inference codeCodeClass IIOSI-approved licence
II.4Evaluation codeCodeClass IIOSI-approved licence
II.5Evaluation dataDataClass IIPreferred CDLA-Permissive-2.0; CC-BY-4.0 and open data licences accepted
II.6Supporting libraries and toolsCodeClass IIOSI-approved licence
I.2Research paperDocumentationClass IPreferred CC-BY-4.0; open content licences accepted
I.3DatasetsDataClass IPreferred CDLA-Permissive-2.0 or CC-BY-4.0; any licence accepted, including none
I.4Data preprocessing codeCodeClass IOSI-approved licence
I.5Intermediate model parameters (checkpoints, optimizer states, log files)DataClass IOpen licence; CDLA-Permissive-2.0 suggested
I.6Model metadataDataOptionalPreferred CDLA-Permissive-2.0; CC-BY-4.0 and open data licences accepted

Expanded article table

Components do not have to be separate files. Evaluation results may sit inside the paper, report or model card, and the data card may be combined with the model card.[2] The supporting libraries entry is broad: the specification lists tokenizers, hyperparameter search code, compute infrastructure code, monitoring code, container files and integration code as things that should be included.[2]

Datasets are the one component for which any licence is accepted. The specification requires that the datasets used to develop the model be provided "in the public domain, as copyrighted data, or under any form of license", recommends an open licence, and lets a producer supply links to curated raw datasets online if the preprocessing code comes with them.[2]

Two further files belong to a compliant distribution. Every distribution must contain a LICENSE file, either one file holding the text of every licence used or one per component. The MOF configuration file is a machine-readable record (JSON, YAML or an AI bill of materials format) with three parts: the framework name and version; the release, including its name, version, date, type, architecture, any post-training treatment, the base model, the producer and the qualifying class; and a list of the included components with the SPDX identifier and licence path for each.[2] The paper called this file MOF.JSON and required it; the specification's appendix says it is now only recommended, although one passage of the specification still says the file "MUST be included in any distribution".[2][5]

Classes

Each class contains everything in the class below it.[2]

ClassNameComponents addedWhat the specification says it enables
IIIOpen ModelModel architecture, final parameters, technical report or research paper, evaluation results, model card, data card; sample outputs optionalUnrestricted use, modification and redistribution; building a product or service; fine-tuning and optimization
IIOpen ToolingTraining, validation and testing code; inference code; evaluation code; evaluation data; supporting libraries and toolsUnderstanding the training process; validating benchmark claims; inference optimization
IOpen ScienceResearch paper; datasets; data preprocessing code; intermediate parameters; model metadata optionalEnd-to-end analysis and auditing; reproduction of a similar model; data exploration

Expanded article table

The specification calls the upper two classes "Open Tooling Model" and "Open Science Model"; the paper uses the shorter names.[2][5]

Class III is the entry level. It asks for the model itself, openly licensed, and enough documentation to use and assess it, but no training code and no training data beyond a data card. An open weights release under a licence with use restrictions does not reach it, and neither does a permissively licensed release that omits one of the documentation components.[6][34] Class II adds the code needed to train, run and evaluate the model. Class I adds the training data, the code that prepared it, checkpoints from intermediate stages of training with their log files, and a research paper.[2]

The requirement for supporting libraries and tools has loosened in practice. Specification 1.0 lists the component under Class II without marking it optional, but presentations by Osborne in September 2025 and Le Hors in May 2026 label it optional, and the tool does not report it as missing when it calculates progress toward Class II.[2][17][19][20]

Model Openness Tool

The Model Openness Tool is the reference implementation of the framework. The site at isitopen.ai describes it as a way to produce a "self-service score": it asks 16 questions about a model and places the release in class 1, 2 or 3.[3] The software is published under the MIT licence in the lfai/model_openness_tool repository on GitHub, where its maintainers are listed as Arnaud Le Hors (IBM), Greg MacKenzie and Sachin Varghese.[4]

How a model is assessed

A producer, or anyone else, fills in the evaluation form with the model's details, the licences that cover the distribution as a whole or all components of one type, and, for each component, whether it is included, where it is and what licence it carries. For every included component the tool looks for a licence in order: one attached to the component, then a global licence for its type, then a global licence for the distribution. A component with no licence counts as not open. A licence that is open but of the wrong type passes with a warning.[4] The tool also applies the specification's substitutions: a research paper makes the technical report optional, evaluation results may live in another document, and a model card that carries the data information makes a separate data card unnecessary.[4]

The form returns a result immediately and can be revised and rerun, so a producer can see which missing component or licence change would raise the class.[4][6] Entries are stored as YAML files in the repository and are added by pull request, which a maintainer must approve. The public registry shows only records marked approved: in June 2025 the maintainers hid the records dating from 2024, so the repository holds more model files (273 in November 2026) than the registry lists.[4][16] Since January 2026 the form can open that pull request itself for a user signed in with GitHub.[4][16]

What the badge means

Each registry page shows the class a model qualifies for and its progress toward the next one, lists the components counted and those missing, and offers a badge with Markdown to paste into a GitHub or Hugging Face page. The badge links back to the registry record.[4][19] The badge programme was modelled on the OpenSSF Best Practices Badge and is free.[5]

A badge records the producer's or submitter's claim; nobody audits it before it is issued. The paper states that "The MOF relies on self-reporting and projects are not classified by a central authority", and that misrepresentation "will only harm the reputation of the model producer"; the community is expected to spot misclassified projects and raise them.[5] The registry's listing and model pages carry a notice that the tool is beta software whose data is preliminary and may be inaccurate.[19][21] Asked in May 2026 how submissions were audited, Le Hors said the project relies on the community: "if you lie and you get caught, you'll get a black eye, right?"[6]

Registry contents

On November 17, 2026 the registry listed 88 models, of which 26 qualified for a class.[21]

ClassModels qualifying (November 17, 2026)
I, Open ScienceAmber (LLM360); Aquila-VL-2B (BAAI); four OLMo variants (Allen Institute for AI); Pythia 70M, 2.8B and 12B (EleutherAI)
II, Open ToolingPolyglot-Ko (EleutherAI)
III, Open ModelGPT-NeoX-20B (EleutherAI); 15 IBM Granite models (four Granite code models, ten Granite 4.0 models and Granite 4.1)

Expanded article table

The other 62 entries were unclassified. Some are close: DeepSeek-R1, released under the MIT licence, and OpenAI's gpt-oss-120b, released under Apache 2.0, are each recorded with five of the six Class III components and are missing only a data card.[34][35] Several Mistral models are shown at the same stage, and the list also holds entries for closed models such as GPT-4.[21] Le Hors said in 2026 that many listed models do not reach Class III because they have no open licence at all.[6]

OpenMDW

The Open Model, Data and Weights License Agreement (OpenMDW) came out of attempts to apply the MOF. Matt White wrote in May 2025 that it was "born out of the effort to implement the Model Openness Framework" and that its scope "maps directly" to the framework, so that every element of a model is covered if it is included in the distribution.[9] The problem it addresses is one the MOF creates: a fully compliant release needs a suitable licence for each type of artifact, with LICENSE files to match. OpenMDW is a single permissive licence that a producer can place at the root of a repository to cover the model, code, data and documentation together.[20][22]

OpenMDW was launched in 2025 by the Linux Foundation and the PyTorch Foundation. Version 1.1, released on May 28, 2026, makes one change: the litigation clause now covers copyright suits as well as patent suits.[18][22] NVIDIA said on the same day that it would adopt the licence for future releases of its Cosmos, Isaac GR00T, Ising and Nemotron model families.[18] Osborne's September 2025 presentation of the MOF licence table recommends OpenMDW for model parameters and describes the licence as compatible with every MOF component.[20]

The licence and the classification are independent. OpenMDW does not require any component to be released, and its FAQ says that MOF classification "is entirely separate" and is led by the Generative AI Commons.[22] A release under OpenMDW can therefore sit in any class or in none. The licensing specialist Shuji Sado argued in July 2025 that this leaves room for openwashing, since a vendor could publish a small part of its data under the licence and keep the rest private.[23]

Adoption

The framework was also taken up outside the Linux Foundation. During the drafting of the Open Source AI Definition, the OSI took the MOF's component list as the vocabulary for its checklist; its process page says "The components we adopted are described in the Model Openness Framework developed by the Linux Foundation."[7] In October 2024 LF AI & Data used the framework as the basis for its own definitions of an "open source AI model" and an "open science AI model", and said community model projects hosted by the foundation would be expected to meet the first.[24] Linux Foundation executive director Jim Zemlin described the MOF in 2024 as "a way to help evaluate if a model is open or not open".[25]

Among model producers, use is concentrated in a handful of organizations. IBM's Granite families are registered, and Granite 4.0 model cards on Hugging Face carry the Class III badge.[19][21][26] The model cards for the Allen Institute's OLMo-7B, LLM360's Amber and BAAI's Aquila-VL-2B carry Class I badges.[27][36][37] According to Le Hors, BAAI's first submission for Aquila "completely failed to qualify", and the institute then reworked the release until it did.[6] Vincent Caldeira, Red Hat's chief technology officer for Asia-Pacific, promoted the framework on the LF AI & Data blog in July 2025 as a way to address opacity in the generative AI supply chain.[28]

In policy discussion the MOF is cited as one of several graded approaches. OpenUK's July 2026 report on AI openness groups it with Stanford's Foundation Model Transparency Index and the European Open Source AI Index as "disaggregation" methods that treat openness as a matter of degree.[29] The paper lists easier regulatory oversight among the framework's benefits, and Osborne's 2025 presentation pairs the MOF and OpenMDW with guidance on the EU AI Act's exemptions for models released under free and open source licences.[5][20]

The name is not unique. The analyst firm Forrester has its own "Model Openness Framework", introduced in an April 2026 blog post: a scoring tool with 12 criteria covering reproducibility, usage rights and community momentum. The post does not mention the Linux Foundation framework, and the two are unrelated.[30]

Relationship to the Open Source AI Definition

The two efforts overlapped in time and in people. The MOF paper appeared in March 2024, while the OSI was still drafting; OSAID 1.0 was finalized in October 2024.[1][6] The MOF paper thanks two OSI representatives among its reviewers, and the OSAID checklist states that it is based on that paper.[5][8] They answer different questions. OSAID defines what must be available for a system to be called open source AI. The MOF does not define that term; it describes how complete and how openly licensed a release is.

Model Openness FrameworkOpen Source AI Definition 1.0
StewardLF AI & Data Generative AI Commons[2]Open Source Initiative[31]
FormThree ranked classes; the tool lists everything else as "Unclassified"[2][21]One definition; a system meets it or does not[31][32]
Unit17 named components[2]Three elements: data information, code, parameters[31]
Training dataNot needed for Classes III and II (data card only); datasets themselves required for Class I, under any licence[2]Detailed information about the data, enough for a skilled person to build a substantially equivalent system; the data itself is not required[31]
Training codeRequired from Class II[2]Required[8][31]
LicencesType-specific: OSI-approved for code, open data and open content licences for the rest[2]OSI-approved licences or terms throughout[8]
AssessmentSelf-reported through the MOT; public registry and badges[5]No certification of individual systems by the OSI[32]

Expanded article table

The OSI's checklist describes the relationship as "a bar overlayed on the MOF range of classes".[8] The bar does not line up with a class boundary. The checklist requires data preprocessing code, which the MOF asks for only at Class I, and training code, a Class II component, but treats evaluation code, evaluation data, evaluation results and the model card as optional.[8] On data the two differ at both ends of the ladder. A Class III model needs only a data card, where OSAID asks for data information detailed enough to rebuild an equivalent system. A Class I model must provide the datasets themselves, which OSAID does not demand.[2][31]

In practice the top of the MOF ladder and the OSAID pass list overlap. The models the OSI reported as passing its validation phase were Pythia, OLMo, Amber, CrystalCoder and T5; Pythia, OLMo and Amber variants are also among the MOT's Class I entries.[21][32]

The use of the MOF inside OSAID drew objections during drafting. In June 2024 Amanda Casari, who signed as OSS and AI lead at Google Open Source, wrote on the OSI forum that the framework "is not agnostic or neutral in how it categorizes, groups together, and classifies different kinds of artifacts", that its authors hold a particular view of which intellectual property law applies to each component, and that it "does not align easily or readily with current releasing and sharing practices". Stefano Maffulli of the OSI replied that the checklist was a working document that could change more often than the definition, and that he expected the Linux Foundation to show the framework in production use.[33] The published checklist notes as a known limitation that, being based on the MOF, it is closely tied to generative AI and not general enough for all machine learning.[8]

Advocates of graded approaches see the difference as the MOF's advantage. OpenUK's 2026 report says the OSAID "did not gain support" and calls disaggregation the "likely better approach".[29] LF AI & Data, for its part, described its own definitions in October 2024 as "much narrower in scope" than the OSI's work.[24]

Criticism and limitations

The framework does not verify entries. Registry data comes from whoever submitted the YAML file, a maintainer approves the pull request, and the only sanction for a false claim is reputational.[4][5][6] The registry describes its own data as preliminary and possibly inaccurate.[21]

Few models hold a class. On November 17, 2026, about two and a half years after the framework was announced, 26 models from five producers did so in the registry.[10][21] Osborne's 2025 presentation puts the scale of the problem at roughly 5,000 models uploaded to the Hugging Face Hub each day.[20] Entries are still written by hand: Le Hors said attempts to generate them with language models had not succeeded, partly because model cards are unstructured, and that "We haven't had anybody really committed to this". He also said, in May 2026, that OpenMDW, the licence meant to simplify compliance, was "not generally used yet".[6] Use has grown since: on November 17, 2026 the Hugging Face Hub listed about 500 model repositories under the openmdw-1.1 tag, among them NVIDIA's Alpamayo 2 and Nemotron releases.[38]

One of the paper's authors has listed the framework's practical problems. Osborne's presentation says the MOF is complicated for producers, who must attach a content-appropriate licence to each artifact and keep LICENSE files in several places, and hard for consumers, most of whom do not understand what the licences imply.[20] Casari's objection during the OSAID drafting, that the framework did not match how developers release models, makes a similar point from outside the project.[33]

The documents also disagree in places. The paper and specification count 17 components, while the tool's site and README refer to 16 questions and 16 components, and the status of supporting libraries and tools differs between the specification and later presentations.[2][3][4][17]

The paper acknowledges limits of scope. The framework was designed for deep learning and does not carry over fully to reinforcement learning, a three-class rubric may miss nuance, and a class says nothing about what a model does. AI safety, bias, performance testing, red-teaming, security, privacy, model serving and provenance are declared out of scope.[2][5]

Training data is the least settled part. The paper states that the framework "does not address the use of copyrighted materials in training data".[5] Class I requires datasets to be provided but accepts them under any licence, so the top class does not by itself mean the data can be reused.[2]

Continue reading

More to read

Topics:AI Policy & RegulationModel EvaluationOpen Source AI

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Matt White, Ibrahim Haddad, Cailean Osborne, Xiao-Yang Yanglet Liu, Ahmed Abdelmonsef, Sachin Varghese, Arnaud Le Hors. "The Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency, and Usability in Artificial Intelligence" (abstract and submission history). arXiv:2403.13784, March 20, 2024; v6 October 18, 2024. arxiv.org/...2403.13784
  2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27LF AI & Data Generative AI Commons. "The Model Openness Framework (MOF) Specification", version 1.0. December 17, 2024. lfaidata.foundation/...paper_MOF_Specification.pdf
  3. ^1 ^2 ^3LF AI & Data. "Model Openness Tool" (isitopen.ai). Accessed November 17, 2026. isitopen.ai
  4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12LF AI & Data. "Model Openness Tool" (README and MAINTAINERS). GitHub, lfai/model_openness_tool. Accessed November 17, 2026. github.com/...model_openness_tool
  5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Matt White et al. "The Model Openness Framework" (full text, version 6: sections 3.10, 4.3, 7.4 to 7.7 and 9, Appendix B). arXiv, October 18, 2024. arxiv.org/...2403.13784v6
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Joe Brockmeier. "MOT: a tool to fight openwashing in AI". LWN.net, May 27, 2026. lwn.net/...1073420
  7. ^1 ^2Open Source Initiative. "Open Source AI Process". Accessed November 17, 2026. opensource.org/...process
  8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Open Source Initiative. "Open Source AI Definition: Checklist". Accessed November 17, 2026. opensource.org/...checklist
  9. ^1 ^2 ^3Matt White. "The Open Source Legacy and AI's Licensing Challenge". The Linux Foundation, May 22, 2025. linuxfoundation.org/...and-ais-licensing-challenge
  10. ^1 ^2 ^3LF AI & Data. "Introducing the Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency and Usability in AI". April 17, 2024. lfaidata.foundation/...parency-and-usability-in-ai
  11. ^1 ^2LF AI & Data. "LF AI & Data Releases Beta Version of Model Openness Tool for MOF Application". June 25, 2024. lfaidata.foundation/...ss-tool-for-mof-application
  12. ^1 ^2Nick Vidal. "Highlights from AI_dev Paris". Open Source Initiative, July 3, 2024. opensource.org/...highlights-from-ai_dev-paris
  13. ^Joe Fay. "Linux Foundation tool reveals the truth about 'open' AI models". The Stack, June 20, 2024. thestack.technology/...how-closed-open-ai-still-is
  14. ^1 ^2Linux Foundation Japan. "LF AI & Data「モデル オープンネス フレームワーク (MOF) 仕様」を公開" (announcement of the Japanese version of the MOF specification, in Japanese). April 7, 2025. linuxfoundation.org/...ess-framework-specification
  15. ^Matt White, Cailean Osborne, Xiaoyang Liu, Keyi Wang, Sachin Varghese. "The Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency, and Usability in Artificial Intelligence" (poster, workshop on Regulatable ML). NeurIPS 2025. neurips.cc/...130232
  16. ^1 ^2 ^3 ^4LF AI & Data. "Model Openness Tool: commit history". GitHub, lfai/model_openness_tool. Accessed November 17, 2026. github.com/...main
  17. ^1 ^2 ^3Arnaud Le Hors. "MOT: A Tool To Fight Open-washing in AI" (slides). Open Source Summit North America, May 18, 2026. hosted-files.sched.co/...%20Minneapolis%202026.pdf
  18. ^1 ^2 ^3The Linux Foundation. "Linux Foundation Releases OpenMDW-1.1; NVIDIA Adopts OpenMDW for Cosmos, Isaac GR00T, Ising and Nemotron AI Model Families". May 28, 2026. linuxfoundation.org/...-nemotron-ai-model-families
  19. ^1 ^2 ^3 ^4 ^5Model Openness Tool. "Granite-4.0-H-Small" (registry record). Accessed November 17, 2026. mot.isitopen.ai/...1164
  20. ^1 ^2 ^3 ^4 ^5 ^6Cailean Osborne. "Practical Tools for Facilitating Openness in AI R&D: The Model Openness Framework and OpenMDW License" (slides). SciLifeLab AI Seminar Series, September 26, 2025. scilifelab.se/...ifeLabAISeminarCaileanOsborne.pdf
  21. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Model Openness Tool. "Models" (registry listing). Accessed November 17, 2026. mot.isitopen.ai/models
  22. ^1 ^2 ^3OpenMDW project. "OpenMDW-1.1 License Agreement FAQ". GitHub. github.com/...FAQ-1.1.md
  23. ^Shuji Sado. "Evaluating OpenMDW: A Revolution for Open AI, or a License to Openwash?". July 14, 2025. shujisado.org/...openmdw-license-review
  24. ^1 ^2LF AI & Data. "Embracing the Future of AI with Open Source and Open Science Models". October 25, 2024. lfaidata.foundation/...rce-and-open-science-models
  25. ^ZDNet, republished by the Cloud Native Computing Foundation. "We're a big step closer to defining open source AI - but not everyone is happy". August 23, 2024. cncf.io/...pen-source-ai-but-not-everyone-is-happy
  26. ^IBM Granite. "granite-4.0-h-small" (model card). Hugging Face. Accessed November 17, 2026. huggingface.co/...granite-4.0-h-small
  27. ^Allen Institute for AI. "OLMo-7B" (model card). Hugging Face. Accessed November 17, 2026. huggingface.co/...OLMo-7B
  28. ^Vincent Caldeira. "Building Trust in GenAI: Why We Need the Model Openness Framework". LF AI & Data, July 15, 2025. lfaidata.foundation/...he-model-openness-framework
  29. ^1 ^2OpenUK. "AI Openness: 3 Years on From Llama 2". July 2026. openuk.uk/...AI-Openness-Report-July-2026-1.pdf
  30. ^Mike Gualtieri. "Introducing Forrester's AI Model Openness Framework". Forrester, April 17, 2026. forrester.com/...sters-ai-model-openness-framework
  31. ^1 ^2 ^3 ^4 ^5 ^6Open Source Initiative. "The Open Source AI Definition 1.0". opensource.org/...open-source-ai-definition
  32. ^1 ^2 ^3Open Source Initiative. "Open Source AI Definition: FAQ" ("Which AI systems comply with the Open Source AI Definition?"). Accessed November 17, 2026. opensource.org/...faq
  33. ^1 ^2Amanda Casari and others. "Concerns and feedback on anchoring on the Model Openness Framework". OSI Discuss forum, June 12-14, 2024. discuss.opensource.org/...393
  34. ^1 ^2Model Openness Tool. "DeepSeek-R1" (registry record). Accessed November 17, 2026. mot.isitopen.ai/...DeepSeek-R1
  35. ^Model Openness Tool. "gpt-oss-120b" (registry record). Accessed November 17, 2026. mot.isitopen.ai/...gpt-oss-120b
  36. ^LLM360. "Amber" (model card). Hugging Face. Accessed November 17, 2026. huggingface.co/...Amber
  37. ^Beijing Academy of Artificial Intelligence. "Aquila-VL-2B-llava-qwen" (model card). Hugging Face. Accessed November 17, 2026. huggingface.co/...Aquila-VL-2B-llava-qwen
  38. ^Hugging Face. "Models" (listing filtered by license tag openmdw-1.1; count read on November 17, 2026). huggingface.co/models

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 4,970 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check on November 17, 2026: 168 claims checked against the paper, the specification and the tool registry. Seven corrections applied before publication.

Cite this page: AI Wiki. "Model Openness Framework." aiwiki.ai, updated 17 Nov 2026, fact-checked 17 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/model_openness_framework

Edit article