Author: Vincent Caldeira, CTO of Red Hat in APAC
This post is part two of our series on transparency in the GenAI supply chain. Following our introduction of the Model Openness Framework (MOF), we now dive into licensing—exploring how the OpenMDW license helps simplify legal complexities and ensures openness across all components of AI model releases.
The Role of Licensing: From OSI to OpenMDW
Operationalizing openness according to the MOF requires clarity and intentionality around licensing. This goes beyond simply choosing a license: it involves understanding which licenses are appropriate for the different types of components that make up an AI model release. Model parameters (weights and biases), for example, are fundamentally data, not software code. Applying a traditional open source software license to model weights can be legally ambiguous, as these licenses were not designed with data-specific considerations like privacy, ethics, or data rights in mind. For this reason, the framework recommends using specific open licenses tailored to the type of artifact being shared:
- OSI-approved open source licenses (e.g., Apache 2.0, MIT) are recommended for code components like model architecture, training scripts, and inference code.
- Open data licenses (e.g., CDLA-Permissive-2.0, CC-BY-4.0) are recommended for data components, including model weights (parameters), datasets, evaluation data, and model metadata.
- Creative Commons licenses (e.g., CC-BY-4.0) are recommended for documentation such as technical reports, research papers, model cards, and data cards.
Navigating the legal landscape of an AI model release can become complex when dealing with multiple components—each requiring a different type of license. Recognizing this challenge, the OpenMDW-1.0 license has been introduced as a novel approach. Developed under the Linux Foundation, OpenMDW, which stands for “Open Model, Data and Weights”, is a permissive license specifically crafted for machine-learning models and their related artifacts, collectively termed “Model Materials”. Under OpenMDW, “Model Materials” broadly encompass the machine-learning model itself (architecture and parameters) and all related artifacts provided in the distribution, such as datasets, documentation, and associated software (preprocessing, inference, training code). OpenMDW fills a critical gap by offering a legal framework that properly addresses rights over trained model weights, architectures, and associated data.
The great advantage of OpenMDW is its comprehensive nature: it can do it all. It replaces the original need to include different licenses for code, data, and content that the MOF (Model Openness Framework) recommends. This solves the challenge of managing multiple licenses and the associated complexity of generating a MOF.JSON file, which typically requires model producers to specify what parts of their repository are covered by what licenses. By providing a single license for this diverse set of “Model Materials,” it aims to ensure consistency and clarity across all components in an open AI model release.
OpenMDW grants broad, unrestricted, royalty-free permission to “deal in the Model Materials without restriction, including under all copyright, patent, database, and trade secret rights included or embodied therein”. This expansive grant directly supports transparency in usage by giving users clear downstream usage rights for fine-tuning, modification, and deployment. This empowers users to confidently experiment, build upon, and redeploy models without navigating ambiguous legal terms often associated with applying code licenses to data or other artifact types, fostering a climate of open collaboration.
The only substantive conditions are the preservation of the OpenMDW license text and all original copyright and origin notices upon redistribution, ensuring proper attribution without imposing restrictive copyleft or share-alike requirements that might hinder flexibility. The license also includes patent defense mechanisms, such as a patent-litigation termination clause, to discourage aggressive patent actions by licensees, further protecting the collaborative ecosystem. Importantly, OpenMDW-1.0 explicitly states that outputs generated by using the Model Materials are not subject to the license’s attribution requirements or other restrictions, ensuring freedom of outputs, which is crucial for commercial and research use cases. Furthermore, the license includes strong warranty disclaimers and places responsibility on the user for due diligence regarding their use of the Model Materials. Finally, it should be noted that OpenMDW is also compatible with other licenses, both permissive and restrictive. As a global license, any components not explicitly covered by another license default to OpenMDW, ensuring comprehensive coverage.
While OpenMDW-1.0 provides clear, permissive terms for using included “Model Materials”, it does not mandate the level of completeness defined by the MOF. The best way to understand this relationship is that the MOF is a classification system that evaluates two key aspects: openness (the use of an open license) and completeness (the inclusion of model artifacts). OpenMDW solves the openness aspect by being applicable to all included components. It does not address completeness, as it does not require that any specific components be included: it only ensures that what is included is released under a permissive license. This significantly reduces implementation complexity by allowing a single LICENSE file in the root of the repository.
Ultimately, OpenMDW is explicitly designed to be compatible with the Model Openness Framework and offers a simplified legal approach compared to managing multiple licenses for different component types. Also OpenMDW embraces all of the principles of the OSI’s open source software definition and is a permissive license. However, unlike the OSAID definition which requires particular components to meet the definition, OpenMDW is compatible with any number of components and puts no restrictions or requirements on model producers or downstream users. This approach has the potential to not only streamline the release process for model producers aiming for MOF compliance but also promote the broadest possible transparent usage of released components.
