Downloading the weights of an AI model can make it possible to run it — and sometimes modify it — without using its developer’s service, if the other necessary components are available and the terms of use allow it. But this is not enough to establish how it was made, which data contributed to its training or what its licence permits. To judge whether a model is genuinely open, availability, documentation and freedom of use must be considered separately.
Weights provide access to the model, not to everything that went into making it
Weights are the numerical parameters adjusted during training. They influence how the model turns an input into a response. When a developer makes them available to download, this gives access to an essential component of the model and may allow users to run it on their own infrastructure.
This does not automatically reveal the training code, how the data was processed or the rights associated with redistribution and modification. “Open weights” therefore describes material availability: the files can be accessed. The term alone does not guarantee that users can study and reuse the system without restrictions.
This distinction matters in practice. An organisation that wants to host a model itself may find the weights sufficient for some uses, subject to the licence and technical requirements. A team that wants to understand the model’s behaviour, modify it in depth or rebuild a comparable system needs other components: documentation, code and information about the data.
What the OSI definition requires
The Open Source Initiative (OSI) published version 1.0 of the Open Source AI Definition, or OSAID, in 2024. This text is not a universal law or an automatic certification. It is a framework that applies the central freedoms of open source to AI: to use the system, study how it works, modify it and share it, without having to ask permission for each use.
To make these freedoms effective, the OSAID requires the “preferred form” for modifying a machine-learning system. This includes the parameters, including the weights, as well as the complete code used to process the data, train and run the system. It also includes sufficiently detailed information about the data for a qualified person to build a substantially equivalent system.
This requirement does not mean that all raw data must be published. The definition calls, in particular, for a complete description of the data used, including data that cannot be shared, its provenance, selection, labelling and processing. It also provides for a list of publicly accessible training data and guidance on how to obtain it. The distinction matters: documenting the materials and process does not necessarily mean distributing every source file.
The OSI specifies that these criteria also apply to models and weights presented as open source. A download can therefore offer a genuine opportunity to experiment while still falling short of this definition if the code or information about the data is missing. To assess an announcement, the documentation actually provided should be compared with these requirements, rather than relying on the label chosen by the developer.
An open licence does not settle every question
The licence determines permissions and terms of use. Permissive licences such as MIT or Apache 2.0 generally allow a wide range of uses, but in themselves they do not say whether the data and training recipe are documented. Conversely, the availability of weights does not guarantee complete freedom: a licence may restrict certain uses or impose specific conditions.
The specific model must therefore be examined, rather than inferring its rights from the reputation of its developer or the name of a model family. The same company may offer closed models, models with weights accessible under their own terms, and others under a more permissive licence. The licence for the downloaded version, any restrictions and the rules on redistribution must be checked separately.
For professional use, this is not an abstract legal detail. It may determine whether the model can be integrated into a commercial service, redistributed to clients or adapted for an internal need. However, it does not answer every question about reliability: an open licence proves neither the quality of the responses nor the absence of bias or errors.
A checklist before choosing
When faced with a model presented as open, three checks provide an initial guide. First, are the weights actually accessible, and can the model be run under the stated conditions? Next, which licence specifically covers this version, and what does it allow in terms of commercial use, modification and redistribution? Finally, does the project publish the code and sufficiently precise information about the data and training to allow independent scrutiny or substantial reproduction?
These questions distinguish needs that are often confused. To test a model or host it internally, the weights and a compatible licence may be enough. To audit it, modify it or understand how the data shaped its responses, access to the weights is no substitute for documentation. And to assess its performance, tests suited to the intended task are still needed: openness is not a quality score.
The term “open source” should therefore not be used as shorthand for “downloadable”, any more than it guarantees that a system is transparent in all its effects. The useful question is more concrete: which components are accessible, under what conditions, and what information is available to genuinely understand or modify the model?
References
Computer code on a screen. Stock photograph; this is not the training code of a model discussed in this article.
Photograph: Martin Vorel — Wikimedia Commons — CC BY-SA 4.0.
