Open Source AI Definition (OSAID)

What it is

The Open Source AI Definition (OSAID) is a standard for when a machine-learning model may be called open source. The Open Source Initiative (OSI) published version 1.0, the first stable version, on 28 October 2024 at the All Things Open conference^[1]. Its purpose is to give the phrase open source AI a fixed meaning that labels and licences can be checked against.

What a model has to disclose

To qualify, a model has to provide enough information about its design that a person could substantially recreate it. It also has to disclose pertinent details about its training data: where the data came from, how it was processed, and how it can be obtained or licensed^[2]. Maffulli described the intent as letting a user fully understand how a model has been built, with access to all the components, including the complete code used for training and for data filtering^[2].

In other words, open source here does not mean only that the weights can be downloaded. It means the process that produced the model can be inspected and repeated by someone else.

The freedoms

The definition lays out usage rights: the freedom to use the model for any purpose, to modify it without permission, and to build on top of it^[2]. These rights are anchored in the Four Essential Freedoms of the classic Open Source Definition^[1], the older document written for software, where the source is program text. Carried over to models, the material that has to be open becomes a mixture of code, data and the training process.

Who checks

OSI has no enforcement mechanisms. It intends instead to flag models that are described as open source but fall short, hoping that the community will correct misuse of the term^[2]. OSI also established a committee responsible for monitoring how the definition is applied and for proposing amendments for future versions^[2].

Where models actually stand

By these criteria, few models qualify. Meta mandates that platforms with more than 700 million monthly active users request a special licence for Llama; Stability AI requires businesses with more than $1 million in revenue to obtain an enterprise licence; and Mistral's licence bars certain commercial uses^[2]. Each of these terms runs against the freedom to use a model for any purpose.

Criticism

Critics make two main points. Luca Antiga, CTO of Lightning AI, argues that the definition does not go far enough: a model can meet every OSAID requirement even when its training data is not freely available, which leaves a licensing hole^[2]. A study by researchers at the Signal Foundation, the AI Now Institute and Carnegie Mellon found that many open source models are basically open source in name only, with training data kept secret and the compute out of reach^[2].

Why it matters

What the definition provides is a common reference point rather than a rule that can be enforced. Anyone releasing a model as open source can now be compared against a written list of requirements, item by item. Because compliance rests on disclosure and community pressure, how much the label means in practice depends on whether it is used accurately.