Open Weights: not quite what you’ve been told – Open Source Initiative

Learn more about OSI’s mission, history, people, and programs.
Explore the list of OSI-Approved licenses and learn more about the license review process.
Learn more about Open Source AI, how do we define it, and its benefits to society.
Advancing Open Source through policy, regulation, and standards.
Supporting research, education, and knowledge about Open Source.
Connecting the people and organizations that build and support Open Source.
Explore the latest news, blog posts, newsletters, press mentions, and events from the OSI.
Learn more about OSI’s mission, history, people, and programs.
Explore the list of OSI-Approved licenses and learn more about the license review process.
Learn more about Open Source AI, how do we define it, and its benefits to society.
Advancing Open Source through policy, regulation, and standards.
Supporting research, education, and knowledge about Open Source.
Connecting the people and organizations that build and support Open Source.
Explore the latest news, blog posts, newsletters, press mentions, and events from the OSI.
In the rapidly evolving landscape of artificial intelligence, Open Weights have emerged as a buzzword indicating incremental progress in AI transparency. By sharing the final parameters of a trained model, developers offer some insight into how a neural network operates. However, these weights reveal only a fraction of the information required for full accountability. While Open Weights represent a milestone in opening up AI systems, they still stop short of delivering the level of transparency many researchers and regulators deem essential.
Open Weights refer to the final weights and biases of a trained neural network. These values, once locked in, determine how the model interprets input data and generates outputs. When AI developers share these parameters under an OSI Approved License, they empower others to fine-tune, adapt, or deploy the model for their own projects.
However, Open Weights differ significantly from Open Source AI because they do not include:
By withholding these critical elements, developers only provide a glimpse into the final state of the model, making it difficult for others to replicate, audit, or deeply understand the training process.
Far from it. Over the past decade, AI practitioners have experimented with different ways of sharing or withholding information, often balancing trade secrets with mounting calls for AI transparency. The renewed interest in Open Weights arose in response to regulatory scrutiny and growing awareness that completely opaque systems can embed biases and discriminatory behaviors.
In 2023, Heather Meeker, a recognized expert in Open Source licensing, published an Open Weights Definition that formalized many of these conversations. Her work clarifies permissible usage and distribution of final model parameters while highlighting a gap: the full process behind model creation still remains undisclosed. As discussions on Open Source AI intensify, so does the debate around whether Open Weights alone can deliver the transparency needed for ethical and responsible AI.
While Open Weights stand out as more transparent than purely proprietary AI, they still lack several key elements of Open Source AI.
1. Lack of reproducibility
Reproducibility is critical in scientific and technological progress. Without training code or intermediate checkpoints, researchers and auditors cannot replicate the model’s development process. This gap hinders efforts to identify when and where biases might have been introduced, making it nearly impossible to rectify errors or vulnerabilities.
The phrase “garbage in, garbage out” applies strongly to AI. If the training data is not representative or ethically sourced, the model’s outputs can exhibit harmful biases. However, Open Weights often do not clarify how the dataset was constructed or cleaned. This oversight leaves a significant blind spot, preventing anyone outside the original development team from fully assessing the dataset’s quality or diversity.
3. Regulatory hurdles
Governments worldwide are formulating policies that mandate higher standards of transparency in AI, especially for systems deployed in sensitive areas such as finance, healthcare, and public administration. Disclosing only the final weights may not meet these emerging regulations, as the lack of training code or dataset details could violate requirements for fairness, privacy, or explainability.
4. Limited community collaboration
One of the core strengths of Open Source AI lies in the collaborative potential it unlocks. When the entire pipeline—training scripts, dataset composition, and intermediate checkpoints—is openly available, a global community can work together to improve the model, fix bugs, or address ethical concerns. By contrast, Open Weights significantly reduce these possibilities, limiting meaningful contributions to superficial fine-tuning rather than in-depth improvements.
Following the same idea behind open source software, an Open Source AI is made available under terms that grant users the following freedoms:
A fundamental precondition to exercise these freedoms is having access to the preferred form needed to make modifications, and the practical means to use it. Open Weights alone fall short of this because they do not provide the underlying training process, code, or comprehensive data details required for full-fledged use, study, modification, and sharing.
To better understand why Open Weights and Open Source AI differ so drastically, consider the following comparison:
Clearly, Open Weights mark a notable advancement over fully proprietary solutions by offering the final model parameters. However, Open Source AI goes further by unlocking the entire development process. This holistic openness enables complete reproducibility, thorough bias audits, and robust community-driven improvements.
* When legally allowed. See Open Source AI Definition FAQ.
A model’s fairness depends he