Author: Vincent Caldeira, CTO of Red Hat in APAC
This post is part one of a multi-part series exploring how the LF AI & Data community is operationalizing transparency in the Generative AI (GenAI) model supply chain. In this installment, we unpack why transparency is vital for responsible AI and introduce the Model Openness Framework (MOF) as a foundational tool for enabling trust.
The Generative AI (GenAI) revolution has ushered in a new wave of innovation, but also introduced a new set of risks. As large language models and multimodal systems become increasingly foundational to digital infrastructure for both enterprises and in the public sector, the question of how they are built, shared, and deployed becomes not just technical, but also ethical and legal becomes even more paramount. In particular, the GenAI model supply chain remains opaque and fragmented, making it difficult to verify claims, reproduce results, or assess the safety and fairness of models that increasingly influence both society and business concerns.
Why Transparency in the GenAI Supply Chain Matters
Transparency is the bedrock of trust. Without transparency, models remain unverifiable black boxes, creating barriers to reproducibility, accountability, and collaboration. It becomes hard to explain models’ internal logic, ensure fairness, audit their behavior, or build upon existing work. Worse, the practice of “openwashing”, representing GenAI models as open source while withholding critical components like training data or fine-tuning scripts, has actually undermined both user confidence and responsible innovation, contributing to the confusion surrounding truly open models.
Governments, regulators, and researchers are sounding the alarm. Without a clear framework for model transparency, it becomes nearly impossible to trace provenance, assess risks, or ensure compliance with emerging AI regulations and frameworks such as the EU AI Act] or the NIST AI Risk Management Framework. Transparency has become vital to facilitate accountability, sovereignty, and collective understanding.
Enter the Model Openness Framework (MOF)
To address these critical challenges head-on, the Model Openness Framework (view MOF specification here), developed by the LF AI & Data Generative AI Commons, defines a tiered classification system for AI models based on the completeness and openness of released components. It’s a core mechanism to evaluate and classify machine learning models, aiming to establish completeness and openness as core tenets of responsible AI R&D, provide practical guidance, promote reproducibility, transparency, and usability, and accelerate progress through open collaboration.
The MOF defines three classes, representing different levels of transparency requirements:
- Class I (Open Science Model): The highest tier, demanding full transparency. This includes releasing training data, all model weights (including intermediate checkpoints), training and inference code, and comprehensive documentation. This ‘comprehensive documentation’ encompasses in-depth research papers that detail the model’s complete methodology, results, and analysis, allowing experts to understand its scientific foundations. Crucially, it also includes detailed model cards and data cards. A model card acts like a ‘nutrition label’ for the AI model itself, providing essential information about its purpose, intended uses, performance metrics (how well it performs), known limitations, and potential risks, helping anyone understand what the model does and its expected behavior. Similarly, a data card provides detailed summary statistics and key information about the dataset(s) used to train the model, describing its composition, how it was collected, and any identified biases or privacy considerations, which is vital for assessing the data’s quality and suitability. Together, this extensive documentation provides the full context necessary for end-to-end analysis, auditing, and reproduction of the AI model, truly embodying the principles of open science.
- Class II (Open Tooling Model): An intermediate tier where key tooling is open. Code (training, inference, evaluation) and model weights are available, but proprietary or sensitive training data may be restricted. Evaluation data and supporting libraries are also provided enabling understanding of the training process, validation of benchmark claims, and inference optimizations.
- Class III (Open Model): The entry point for openness. This involves the partial release of components which typically includes the model architecture, final model weights, and basic documentation, like a technical report, evaluation results, a model card, and a data card. This class provides for unrestricted usage in products and services, along with fine-tuning and model optimizations.
Each class defines a minimum set of components—17 in total across architecture, data, code, and documentation—that must be included and released under appropriate open licenses to qualify for that tier. Critically, the MOF is license-aware: true openness requires not just the publication of artifacts, but the legal permission for others to freely access, use, modify, and redistribute them.
