OpenMDW

OpenMDW (Open Model, Data and Weights License) is a permissive license agreement developed under the Linux Foundation specifically for machine learning models and their related artifacts.^[1][2] Unlike traditional software licenses, which address only copyright in code, OpenMDW grants rights under "all copyright, patent, database, and trade secret rights" in a single instrument covering what the license calls "Model Materials": one or more machine learning models (architecture and parameters) together with all related artifacts, such as datasets, documentation, and software, that a publisher chooses to distribute under it.^[3] The license text was first published as OpenMDW-1.0 in April 2025, and a minor revision, OpenMDW-1.1, followed in May 2026.^[2][4][9] OpenMDW gained prominence through adopters including NVIDIA, which announced in May 2026 that it would apply OpenMDW-1.1 across future releases of its Cosmos, Isaac GR00T, Ising, and Nemotron open model families,^[9] and which released its Alpamayo 2 Super autonomous-vehicle reasoning model under OpenMDW-1.1 on August 4, 2026.^[10][11]

Key facts

Scroll sideways for more →

History and purpose

OpenMDW grew out of the Model Openness Framework (MOF), a classification system introduced by Linux Foundation AI and Data communities for describing how complete and open an AI model release is.^[1][2] Existing practice for "open" model releases typically combined a software license such as the Apache License or MIT License for code with separate data licenses for datasets, or used custom model licenses with usage restrictions; the Linux Foundation argued that none of these were designed for the mixed composition of a model distribution, in which trained weights may not clearly be covered by copyright at all.^[1][6][16]

The license was drafted by a group of attorneys and open source specialists from the Linux Foundation and member companies. The repository's credited contributors are Steve Winslow (Linux Foundation), Matt White (PyTorch Foundation/Linux Foundation), David Platz (Amazon), Jonathan Torres (Meta), John McBroom and Robert Shatto (IBM), David Rudin (Microsoft), Michael Dolan (Linux Foundation), and Meredith Price and George Chellapa (NVIDIA).^[2] The initial license text was committed to the public OpenMDW GitHub repository on April 19, 2025,^[2] and the launch was covered in trade press by mid-May 2025.^[7] Matt White, executive director of the PyTorch Foundation, presented OpenMDW in a May 22, 2025 Linux Foundation blog post as the licensing complement to the MOF, standardizing terms for models the way Apache-2.0 and MIT did for software.^[1] The Linux Foundation published a fuller rationale on the Hugging Face blog in July 2025.^[6]

OpenMDW-1.1, announced by the Linux Foundation on May 28, 2026, is a minor release containing a single substantive change from version 1.0: the litigation-termination clause was broadened from covering patent lawsuits to covering "patent or copyright" lawsuits.^[4][9] The revision was added to the GitHub repository on May 19, 2026.^[2]

License terms

OpenMDW is deliberately short, structured similarly to permissive software licenses. Its main provisions are:^[3][4]

• Broad grant. Permission is granted free of charge "to deal in the Model Materials without restriction, including under all copyright, patent, database, and trade secret rights included or embodied therein." There are no field-of-use, royalty, or geographic restrictions, and no copyleft or share-alike obligations.
• Notice preservation. Anyone who redistributes any portion of the Model Materials must include a copy of the license agreement and retain applicable copyright and origin notices. This is the license's only substantive condition.
• Litigation termination. If a licensee files or maintains a lawsuit asserting that the Model Materials directly or indirectly infringe a patent or copyright (other than defensively), all rights under the license terminate. The OpenMDW FAQ describes this as comparable in intent to Apache-2.0's defensive patent termination, though broader in scope.
• Output freedom. The agreement imposes no restrictions or obligations on "any use, modification, or sharing of any outputs generated by using the Model Materials," an explicit answer to a question most software licenses do not address.
• Disclaimers. The materials are provided "AS IS" with all warranties disclaimed, and users are solely responsible for clearing third-party rights that may apply to the materials or their use.

The license does not require completeness: a publisher may release only weights, or weights plus code, and OpenMDW covers whatever is actually provided under it. Classification of how complete a release is remains the role of the separate Model Openness Framework.^[2][4][5] Version 1.0 of the license is on the SPDX License List under the identifier OpenMDW-1.0, which is marked as neither OSI-approved nor FSF-libre; as of May 2026 the project's FAQ stated that version 1.1 had not yet been submitted to SPDX.^[4][8]

Adoption

Adoption was initially slow. As late as May 27, 2026, IBM engineer Arnaud Le Hors, speaking about openwashing in AI at the Linux Foundation's Open Source Summit North America, described OpenMDW as a license meant to cover models and all their artifacts "but it's not generally used yet."^[15]

The following day, the Linux Foundation announced OpenMDW-1.1 together with NVIDIA's commitment to adopt the license across future releases of four open model families: Cosmos (world models), Isaac GR00T (robotics), Ising (quantum computing), and Nemotron.^[9] Kari Briski, NVIDIA vice president of generative AI, said in the announcement that "broad access to models, data, and tools is essential to accelerating breakthroughs."^[9]

The most prominent release under the license to date is NVIDIA Alpamayo 2 Super, a reasoning vision-language-action model for autonomous vehicles that NVIDIA made generally available for commercial use under OpenMDW-1.1 on August 4, 2026.^[10][12] NVIDIA CEO Jensen Huang wrote that the company was "releasing it for commercial use under OpenMDW-1.1 so teams can inspect it, fine-tune it and deploy it."^[11] As of early August 2026, NVIDIA's Hugging Face account listed eight model repositories under the OpenMDW-1.1 tag, including Alpamayo 2 Super, the earlier Alpamayo-R1-10B and Alpamayo-1.5-10B research models, an Ising quantum error correction decoder, and the NemotronLabs VoiceChat speech model.^[12] Outside NVIDIA, the AI software development company poolside distributes its Laguna S 2.1 and Laguna XS 2.1 open-weight coding models, published on Hugging Face in June and July 2026, under OpenMDW-1.1.^[13]

Relationship to the open source AI debate

OpenMDW takes a different approach from the Open Source Initiative's Open Source AI Definition (OSAID), which specifies what components must be released for a system to qualify as open source AI. OpenMDW instead defines the rights conveyed for whatever a publisher chooses to share, without mandating that training data or any other component be included.^[2][5][16]

Commentators have assessed the license from both directions. Japanese open source licensing specialist Shuji Sado concluded in a July 2025 review that OpenMDW is a well-constructed application of MIT and Apache-2.0 concepts to AI systems and likely satisfies the criteria of the Open Source Definition, but warned that, because it does not require completeness, it could enable "openwashing": marketing a model as open source AI while withholding critical training data.^[14] Writing for diginomica in June 2026, Alyx MacQueen reported the Linux Foundation's argument for the license over Apache-2.0 applied to weights: because copyright may not attach to machine-generated weight files, an Apache license on weights may convey little, whereas OpenMDW affirmatively grants patent, database, and trade secret rights alongside copyright. Michael Dolan, the foundation's senior vice president of legal and strategic programs, said the license ensures users receive "any of the copyright, patent, database, trade secret rights that may have been embedded" in the materials.^[16] For distributions using open weights without full training data, OpenMDW licensing therefore clarifies the legal terms of what is released without settling the definitional question of whether the release qualifies as open source AI.^[5][14]

See also

• Linux Foundation
• Open source AI
• Open weights
• Apache License
• NVIDIA Alpamayo 2 Super

References

1. ^Matt White, The Linux Foundation. "The Open Source Legacy and AI's Licensing Challenge." May 22, 2025. linuxfoundation.org/...and-ais-licensing-challenge
2. ^OpenMDW project. "OpenMDW License" (GitHub repository and README). Accessed August 5, 2026. github.com/...OpenMDW
3. ^OpenMDW project. "OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)" (license text). github.com/...LICENSE.OpenMDW-1.1
4. ^OpenMDW project. "OpenMDW-1.1 License Agreement FAQ." github.com/...FAQ-1.1.md
5. ^LF AI & Data Foundation. "Simplifying AI Model Licensing with OpenMDW." July 22, 2025. lfaidata.foundation/...odel-licensing-with-openmdw
6. ^The Linux Foundation (Hugging Face blog). "Why We Built the OpenMDW License: A Comprehensive License for ML Models." July 2, 2025. huggingface.co/...openmdw
7. ^The New Stack. "OpenMDW Is a New Permissive License for AI Models." May 18, 2025. thenewstack.io/...permissive-license-for-ai-models
8. ^Software Package Data Exchange (SPDX). "OpenMDW License Agreement v1.0." spdx.org/...OpenMDW-1.0
9. ^The 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
10. ^NVIDIA Blog. "Alpamayo 2 Super Open Model Now Available." August 4, 2026. blogs.nvidia.com/...super-open-model-now-available
11. ^Jensen Huang (@JensenHuang) on X. August 4, 2026. x.com/...2084656303046332747
12. ^NVIDIA. "Alpamayo2-Super" (Hugging Face model page, license: OpenMDW-1.1). huggingface.co/...Alpamayo2-Super
13. ^poolside. "Laguna-S-2.1" (Hugging Face model page, license: OpenMDW-1.1). huggingface.co/...Laguna-S-2.1
14. ^Shuji Sado. "Evaluating OpenMDW: A Revolution for Open AI, or a License to Openwash?" July 14, 2025. shujisado.org/...openmdw-license-review
15. ^Joe Brockmeier, LWN.net. "MOT: a tool to fight openwashing in AI." May 27, 2026. lwn.net/...1073420
16. ^Alyx MacQueen, diginomica. "What OpenMDW 1.1 guarantees an enterprise that an Apache license on model weights can't." June 3, 2026. diginomica.com/...pache-license-model-weights-cant