These terms are used interchangeably in almost all coverage, including by people who know better. They mean different things, and during 2026 the difference became the practical centre of a sanctions argument.
The definitions
Open weights means the trained parameters are published. You can download them, run the model on your own hardware, fine-tune it, and inspect it. That is a meaningful freedom and it is what most people actually want.
Open source, in the sense the term has carried in software for decades, means something stricter: freedom to use, study, modify and redistribute, for any purpose, without field-of-use restrictions. The Open Source Initiative published an Open Source AI Definition in 2024 extending this to models, and it requires meaningful access to the components needed to reproduce the system, not just the finished artefact.
Why most “open source” models are not
Three things typically disqualify a model release:
- Use restrictions. Many popular licences prohibit specific applications or impose conditions above a user threshold. Any field-of-use restriction is disqualifying under an open-source definition, however reasonable the restriction seems.
- No training data. Weights without data or training code are not reproducible. You cannot study how the system came to behave as it does, only that it does. In software terms you have been given the binary, not the source.
- Licences that are not licences. Several model releases use bespoke agreements with acceptable-use policies attached, which function more like terms of service than like open-source licences.
None of this makes open-weights releases bad. They are enormously valuable. It makes the label wrong, and the label is doing real work in policy discussions.
Why it mattered in 2026
The distinction stopped being pedantic when Moonshot published the Kimi K3 weights four days after the White House accused the company of building that model through unauthorised distillation.
Once weights are public, they are public permanently. Enforcement mechanisms aimed at a company, such as Entity List designation or restrictions on commercial operations, do not reach a model that has already been distributed. Any capability in it is in circulation for good.
This is the asymmetry that makes open-weights releases strategically significant in a way that open-source licensing is not. The licence can be restrictive and it barely matters. The distribution is what cannot be undone.
What open weights gives you, and what it does not
You get: local and air-gapped deployment, no per-token cost, no vendor dependency, no data leaving your infrastructure, the ability to fine-tune freely, and the ability to keep running the model after the publisher loses interest.
You do not get: reproducibility, insight into training data provenance, any guarantee about what the model absorbed during training, or, usually, permission to do absolutely anything you want with it.
That last point deserves attention. A model can be freely downloadable and still carry a licence that prohibits your specific commercial use. Read it.
How to check what you actually have
- Find the licence, not the announcement. Marketing language and licence terms frequently disagree.
- Look for field-of-use restrictions. Any clause limiting what you may build immediately rules out open source.
- Look for scale thresholds. Conditions that trigger above a certain user count are common and easy to miss.
- Ask whether training data is disclosed. Almost never. This is the clearest signal that you have weights rather than source.
- Check redistribution terms if you intend to ship the model onward, including inside a product.
The terminology is not going to be fixed
“Open source AI” is established in common usage and will keep being applied to open-weights releases regardless of definitions. That is worth accepting rather than fighting.
What is worth insisting on is precision where consequences attach. In a licence review, a compliance discussion, or a policy debate about what enforcement can reach, the difference between you may run this and you may do anything with this is the whole question.
Common questions
Is an open-weights model the same as open source?
No. Open weights means you can download and run the parameters. Open source, in the sense the Open Source Initiative defines it, requires freedom to use, study, modify and redistribute for any purpose, with no field-of-use restrictions. Many popular models fail that test because their licences restrict acceptable use or impose scale-based conditions.
Why does the difference matter?
Practically, because restrictive licences can prohibit your use case, and legally, because once weights are published they cannot be recalled. Distribution is irreversible in a way that API access is not.
Can you tell how a model was trained from its weights?
Not usefully. Weights are the output of training, not a record of it. Without the training data and code, an open-weights release is not reproducible, which is a large part of why it does not meet an open-source definition.