How the Open Source Initiative's 1.0 Definition Excludes the Most Downloaded Open-Weight AI Models

On October 28, 2024, the Open Source Initiative published version 1.0 of its Open Source AI Definition, establishing a standard that immediately disqualified the vast majority of the artificial intelligence industry's publicly available models. The document, drafted after two years of consultation, dictates that an AI system can only be called "open source" if a developer provides the full training code, the model parameters, and sufficiently detailed information about the training data for a skilled person to recreate the system.[1]

That data requirement strikes at the core of how modern artificial intelligence is built and distributed. While companies like Meta and DeepSeek have aggressively released the numerical weights of their models—allowing anyone to download and run them—they uniformly withhold the underlying datasets and the specific code used to filter that data.[2][5][6]

The distinction between "open-source" and "open-weight" is not merely semantic; it defines what a downstream user can actually do with the technology. Model weights are the central parameters learned during the training process, determining how a neural network interprets new data and generates predictions.[2][6]

Releasing these weights allows independent developers to fine-tune a model for specific tasks, such as medical diagnosis or legal analysis, at a fraction of the cost of training a system from scratch. However, without the original training data, researchers cannot fully audit the model for embedded biases, verify the provenance of copyrighted material, or understand exactly how the system arrived at its baseline capabilities.[2][6]

As IBM researchers noted in a May 2025 technical assessment, "Open weights are publicly shared and typically available under open-source licenses, providing a peek into a deep learning model's final state. And while they signify gradual advancement toward transparency in AI, open weights still don't offer the full picture that open-source AI does." The assessment highlighted that models ranging from 7 billion parameters to 40 billion parameters are frequently mislabeled in public discourse.[2]

The Open Source Initiative's 1.0 framework attempts to bridge this gap by demanding what legal scholars term a "take it or leave it" approach. To claim the open-source label, a developer must grant users four fundamental freedoms: to use the system for any purpose, to study how it works, to modify it, and to share it.[1][3]

Crucially, the OSI specifies that exercising these freedoms requires access to the "preferred form to make modifications." In traditional software, this means the source code. In machine learning, the OSI argues, this preferred form must include the data information.[1]

The definition explicitly requires "the complete description of all data used for training, including (if used) of unshareable data, disclosing the provenance of the data, its scope and characteristics, how the data was obtained and selected, the labeling procedures, and data processing and filtering methodologies."[1]

By this standard, the most widely adopted models in the developer ecosystem operate in a gray area. A July 2025 report by the Centre for Future Generations (CFG) analyzed the distribution of these systems, noting that the Hugging Face repository hosts more than 200,000 text-generation models.[4][7]

The CFG researchers found that five of the top ten most downloaded text generation models on the platform are general-purpose advanced AI models released within the preceding year. None of these leading models provide the comprehensive data provenance and filtering code mandated by the OSI 1.0 definition, despite driving a 98% year-over-year growth in open generative AI projects on GitHub in 2024.[4][7]

This widespread adoption of open-weight models has created a parallel ecosystem where highly capable systems are freely accessible but fundamentally opaque. The CFG report highlights that while open-weight releases democratize access to powerful tools, they also introduce systemic vulnerabilities that traditional open-source software does not face.[4][5]

"When model weights are openly available, individuals can bypass built-in safeguards—a process known as jailbreaking—or fine-tune the model on new data to alter its outputs with minimal cost and effort," the CFG analysis states. Because the weights cannot be recalled once downloaded, any embedded safety mechanisms can be systematically stripped away by motivated actors.[4][6]

To address these risks, governance researchers are pushing to move the debate beyond the binary classification of "open" versus "closed." The CFG advocates for a tiered, safety-anchored approach to model release, where the degree of openness is determined by rigorous risk assessments rather than ideological commitments.[4]

Under a tiered framework, models that exceed specific capability thresholds for dual-use risks—such as assisting in cyberattacks or biological weapon design—would be restricted to verified-access programs, even if their developers wished to release the weights publicly.[4]

The stakes of this definitional battle extend directly into international law. Article 53 of the European Union's AI Act provides specific exemptions from documentation and transparency obligations for AI models released under free and open-source licenses.[3]

The next test of this standard will arrive as regulators in the European Union and the United States finalize their enforcement frameworks. If those bodies adopt the Open Source Initiative's strict criteria, the developers of open-weight models will be forced to choose between disclosing their proprietary training data or accepting the full liability burden of a closed system.[1][3]