Why We Built the OpenMDW License: A Comprehensive License for ML Models
By The Linux Foundation
In the last several years, machine learning has rapidly matured from experimental research into production infrastructure. Yet the legal frameworks that support the open and collaborative development and distribution of machine learning models have lagged behind. Existing licenses, whether permissive open source software licenses like MIT or Apache 2.0, or restrictive community licenses for "open weights" models were never designed for the unique composition, use, and distribution patterns of machine learning models; or were restrictive and imposed exclusions on groups of users or use cases. It was time to address this gap.
That's why, working through the Linux Foundation and in collaboration with Amazon, Meta, IBM, Microsoft, and others, we developed the Open Model, Data & Weights License Agreement (OpenMDW). This license was purpose-built to meet the specific legal and practical needs of sharing machine learning model materials, including large language models (LLMs), openly, responsibly, and effectively.
The Problem: Traditional Software Licenses Don't Fit ML Models
Permissive software licenses like MIT, BSD, and Apache 2.0 have powered decades of innovation in open source software. But machine learning models are not just code. They can include:
- Architectures (in configuration and/or code)
- Trained parameters (weights & biases)
- Metadata and additional configuration files
- Datasets for different stages/types of training
- Software and scripts (data preprocessing scripts, training and inference code)
- Benchmarking code and datasets
- Tokenizers
- Model outputs
- Documentation (technical reports, research paper, model & data cards, manuals)
Besides software, traditional licenses don't explicitly account for these artifacts. As a result, when a model is released under, say, the Apache 2.0 License, it may be unclear:
- Whether the trained weights are licensable under typical understandings applicable to IP rights in software.
- What rights users have over generated outputs or what requirements may apply.
- How patents, trade secrets, and data licenses factor into reuse.
Different types of licenses and agreements are typically used for different types of content. For example, an open source software license like MIT might be used for source code; Creative Commons licenses for documentation; and a data-focused agreement like CDLA-Permissive-2.0 for data. Although it can be useful for model distributors to choose appropriate licenses for different content types, this can also be a complex task to get right for AI/ML specialists who aren't also trying to become legal specialists at the same time.
This lack of clarity has created challenges and complexities for companies, researchers, and developers alike. OpenMDW was designed to fix that.
A License Tailored for Machine Learning
The OpenMDW License Agreement v1.0 provides a clear and unified framework to license all the components of a machine learning model distribution—what we call "Model Materials."
Key Features:
Comprehensive Scope: Covers code, parameters, datasets, and other components necessary to use, reproduce, or modify a model.
Flexible Outputs: Output from the model is free from license obligations, which means you can use and redistribute the outputs without restrictions imposed by the license.
Permissive Terms: Grants broad, royalty-free rights under copyright, patent, database, and trade secret law, requiring retention of notices only when redistributing the Model Materials themselves (and not their outputs).
Patent Safety: Includes a patent termination clause to discourage aggressive litigation while preserving openness.
Compatibility and Dual Licensing: OpenMDW is designed to coexist with other open licenses and supports dual licensing models—providing the flexibility real-world ML deployments need.
Covered but not required: Does not require that any or all components be included in the Model Materials, but does cover what is included and is not subject to other licenses.
Global and Interoperable
From day one, OpenMDW was designed as a global license that supports diverse distribution approaches. It allows model developers to:
- Use third-party components under different open licenses.
- Combine model code and parameters under OpenMDW while releasing datasets under Creative Commons licenses or other compatible terms.
- Enable enterprise use while maintaining openness and transparency.
This flexibility is especially important in ML, where models are often built from modular, distributed components developed by different parties and foundation models are released, then are fine-tuned, quantized or further modified by the community.
Built by and for the Community
OpenMDW is not just a legal document—it's the product of collaboration between leaders in open source software and machine learning. The license was authored through the Linux Foundation, with contributions from legal and technical experts at:
- Amazon
- Meta
- IBM
- Microsoft
- The PyTorch Foundation
- The Linux Foundation
Our shared goal was to create a license that is both legally sound, practically useful for modern ML workflows, and easy for the community to understand.
Why OpenMDW Is the Right Choice for Model Distribution
Whether you're releasing a state-of-the-art LLM or a specialized domain model, OpenMDW provides the clarity, coverage, and flexibility you need:
- You can release model materials with confidence that users understand their rights.
- Downstream users can build, fine-tune, and deploy models without worrying about ambiguous or missing terms.
- Organizations can comply more easily with license terms across the ML stack.
In short, OpenMDW is a license for how machine learning is actually developed, shared, and deployed today. OpenMDW aims to be the de facto standard for permissive licensing of open models and will help provide the community with the confidence needed to know that models licensed under OpenMDW can be modified, studied, and redistributed freely, for any purpose.
Learn More
You can read the full license agreement at openmdw.ai/license and explore the FAQ at openmdw.ai/faq. We encourage model developers, compliance and legal teams, and open source communities to review and adopt this license. It's open by design—and intended to evolve with the ecosystem.
As the AI world continues to move fast, let's make sure our legal infrastructure keeps up. OpenMDW is a step toward that future.
