Model access and licensing concept

Open weights versus open source

Reference

Open weights vs open source: open weights describes model parameters made available to users; the Open Source AI Definition from the Open Source Initiative (OSI) also requires freedoms and materials for modification. These labels answer different questions, and neither tells you every detail about a particular release. Check which files are provided, whether access is gated, and what terms apply to your intended use.

Reviewed 2026-10-09

Two different kinds of openness

A model’s weights are learned parameters used with its architecture to produce outputs. An open-weight release makes those parameters available. The label alone does not tell you whether training code, information about training data, or particular reuse permissions are also provided. OSI’s explanation of open weights describes the distinction.

The Open Source Initiative’s Open Source AI Definition describes freedoms to use, study, modify, and share an AI system, along with access to the preferred form for making modifications. For machine-learning systems, the definition specifies data information, code, and parameters as materials for modification. Read the OSI definition.

What the labels do and don’t tell you

The practical difference is scope. Open weights tells you that parameters have been made available; by itself, it does not establish that training code and data information are provided, or that the release grants the freedoms in OSI’s Open Source AI Definition. That definition also describes freedoms and materials needed to modify machine-learning systems.

Neither label replaces checking a release’s own page and terms. A model card can display license information and link to a custom license. Access can also be gated: Hugging Face says users may need to request access and be logged in before using a gated model. See its model-card documentation and gated-model documentation.

Which distinction matters for your task?

If you mainly want to run a model using released parameters, start by checking whether you can obtain the files. Then read the specific terms for your intended use, and see whether the release includes the code or data information you need. A label about weights alone does not answer those separate questions.

If you need to examine or modify how a system was made, or share a modified system under the freedoms OSI describes, assess the release against the Open Source AI Definition. It sets out freedoms and describes the data information, code, and parameters that make up the preferred form for modifying a machine-learning system.

Common confusions

“The weights are downloadable, so the model meets OSI’s Open Source AI Definition” skips the definition’s broader requirements. The definition covers freedoms to use, study, modify, and share, and identifies materials for making modifications. A release can offer weights without establishing that it meets those requirements. Compare the definition with OSI’s explanation of open weights.

“OSI’s Open Source AI Definition requires publishing every raw training file” is not what the definition says. It specifies detailed information about training data, including information about data that cannot be shared. And an open-weight release is not necessarily ungated: a host may require login or an access request before files are available. Check both the release materials and its access and use terms.

Questions people ask

Are open-weight models open source?

Not automatically. Open weights describes the availability of parameters. Under OSI’s Open Source AI Definition, the term also involves freedoms to use, study, modify, and share, plus materials for making modifications.

What does open source AI mean?

Under OSI’s Open Source AI Definition, it means an AI system is provided with freedoms to use, study, modify, and share, and with the preferred form and materials needed to make modifications.

Can an open-weight model have gated access?

Yes. A host can require users to log in or request access before downloading files. Check the repository’s access instructions as well as the model’s terms.

Does OSI’s Open Source AI Definition require publishing all training data?

It specifies detailed information about training data, including information about data that cannot be shared. It does not say every raw training file must be published.

Learn more

The Open Source AI Definition
Read OSI’s freedoms and its requirements for parameters, code, and data information.
Open Weights: not quite what you’ve been told
See OSI’s explanation of what the weights label does and does not establish.
Model Cards
See how a model card can display license information and link to a custom license.

What changed

Sources

Part of Model access and licensing: Access, weights, and the OSI definition Check terms for your intended use Understand gated and service access