The Model Openness Framework: Promoting Completeness and Openness for Reproducibil...

## Overview Matt White, Executive Director of the PyTorch Foundation and GM of AI at the Linux Foundation, presents the Model Openness Framework (MOF), a collaborative initiative by the Generative AI Commons, Linux Foundation, Columbia University, and University of Oxford to standardize and promote genuine openness in AI model development and release.

## The Problem: Defining "Open" in AI White identifies widespread confusion and misuse of the term "open" in AI. He highlights several key issues: "open washing" where models are marketed as open despite restrictive licenses; models released under licenses with downstream usage restrictions; unclear model lineage when fine-tuned models are re-released under different licenses; and ambiguity around whether traditional open source software licenses (like Apache 2.0) actually cover model weights. Many supposedly open models are actually "source available" rather than truly open, and organizations lack understanding of licensing implications.

## Core Concepts: Openness vs. Completeness The framework disentangles two distinct concepts. Completeness refers to releasing all components involved in model development (data, code, parameters, documentation). Openness is defined more narrowly as a binary property: whether components use OSI-approved open licenses. The framework borrows four freedoms from open source: freedom to use, modify, and distribute (including commercially); accessibility to download components; collaboration and community contribution; and transparency to scrutinize and improve the work.

## Three-Tier Classification System The framework establishes three classes:

**Class 3 (Open Model)**: Minimum viable tier including model architecture, model parameters/weights with optimizer state, technical report with evaluation results, model cards, and data cards—all components under OSI-approved licenses.

**Class 2 (Enhanced Transparency)**: Adds evaluation data and code, training code, inference code, and documentation of libraries used, enabling reproducibility of claimed performance metrics.

**Class 1 (Open Science)**: Most complete tier providing full datasets, data preprocessing code, all model metadata, and intermediate training checkpoints for progressive transparency into the training process. Noted as the aspirational tier to strive toward.

## Implementation: MOF JSON and Tooling The framework introduces a MOF JSON index file that catalogs all released components and their associated open licenses, providing consumers explicit clarity on what they receive and what they can do with it. A web-based tool (planned for AI Dev Paris conference in June) allows model producers to submit their models, select components and licenses, receive a classification, and obtain a badge (modeled on OpenSSF best practices badges) for documentation. Results are recorded in a model openness scoreboard.

## Key Technical Insight: License Ambiguity White emphasizes an unresolved legal gray area: Apache 2.0 and other open source software licenses were designed for code, not model weights. It is legally untested whether these licenses implicitly cover model parameters. This creates risk scenarios where a model released under Apache 2.0 could later face legal challenge if the licensor claims weights were never covered. He provides an example of fine-tuned LLaMA 2 models (released under a restrictive license) being re-released as Apache 2.0, which violates the original license terms and obscures model lineage.

## Benefits and Incentives For model producers, releasing under open licenses with high completeness fosters ecosystem adoption, especially in Enterprise; enables community innovation and feedback; improves model safety through distributed red teaming; and aligns with emerging regulatory scrutiny from agencies like NIST and NTIA. For consumers, the MOF clarifies usage rights, enables safe commercial deployment, permits fine-tuning for domain-specific purposes without data-sharing obligations, and supports reproducibility and collective improvement.

## Scope and Limitations The framework explicitly excludes bias/fairness audits, AI safety assessment, code/model review and auditing, performance validation, legal compliance, and red teaming—delegating these to specialized organizations and existing regulatory frameworks. The MOF focuses narrowly on two outcomes: encouraging release of more components under open licenses and informing the public about what each model actually provides.

## Validation and Adoption White notes initial validation partnerships with Aleph AI and rwkv, who have agreed to evaluation of their model catalogs. The primary adoption challenge is educating both researchers and consumers about licensing implications, and creating market demand for permissive models. He argues that once a few high-performing models adopt fully open licensing, market momentum will shift toward openness by default.

## Takeaways The Model Openness Framework attempts to resolve ambiguity in AI openness claims through transparent classification of model completeness and enforcement via open licenses. It represents a working definition from the open science and open source communities applied to the AI era, with practical tooling for implementation and community oversight via scorecard tracking.