Open Weights vs Open Source: The Distinction That Matters
5 min read · updated August 3, 2026
A model whose weights you can download is not thereby open source. The gap between those two ideas is where the interesting obligations live, and it is the reason a team can build for six months on a model they were never permitted to ship.
Two terms that are not synonyms
Open weights is a statement about availability: the trained parameters are published and you can download them. That is all it says. It says nothing about what you may do with them.
Open source is a statement about permissions, and it has a definition with teeth. The traditional software definition requires, among other things, that the licence not discriminate against any field of endeavour and not restrict what you do with the program. The Open Source Initiative later published a definition aimed at AI systems specifically, which asks additionally for enough information about the training data, the training code and the parameters that a competent party could rebuild something equivalent.
Measured against either, most popular downloadable models are not open source. They carry a bespoke licence with use restrictions attached, which is exactly the kind of clause the field-of-endeavour rule excludes. This is not a scandal — a permissive-with-conditions licence can be a perfectly reasonable thing to build on — but calling it open source causes people to skip reading it, and the conditions are the part that matters.
What actually gets released
Before asking about permissions, ask what was even published. A release can include any subset of:
- Weights. The trained parameters. Almost always present, or there would be nothing to discuss.
- Inference code. How to run them. Frequently present and frequently under a different, more permissive licence than the weights — read both.
- Training code. How they were produced. Sometimes.
- Data, or a description of it. Rarely the data itself; occasionally a documented mixture. This is the single most common gap, and it is the one the OSI’s AI definition presses on hardest.
- Intermediate checkpoints and the recipe. Almost only from research groups whose product is the openness — projects like AI2’s OLMo and EleutherAI’s Pythia exist partly to be the fully documented case.
A useful shorthand: if you cannot in principle reproduce the model, you have open weights. If you can, and the licence does not stop you, you have something that deserves the other name.
The licence families
These are families, not a catalogue. Vendors change licences between releases — the same lab may ship one model under a permissive open-source licence and the next under a bespoke one — so the family tells you what to expect and the LICENSE file in the repository tells you what is true.
| Family | Description |
|---|---|
| Apache-2.0 / MIT | Genuine open-source licences. Commercial use, redistribution and derivatives are all permitted, with no field-of-use restriction and no user-count threshold. Apache-2.0 adds an express patent grant that terminates if you sue over patents, and asks you to preserve notices and state your changes. If a model you want is available under one of these, the licence question is essentially closed. |
| Vendor community licences | Bespoke licences written by the lab that trained the model — Meta’s Llama community licence and Google’s Gemma terms are the archetypes. Broadly permissive commercially, but with conditions: an acceptable-use policy incorporated by reference, a duty to pass the licence and the policy on to anyone you redistribute to, attribution and derivative-naming requirements, and in some cases a scale threshold above which you must go and negotiate separately. Not open source, and usually workable. |
| Research / non-commercial | Research-only weights, sometimes under a vendor research licence and sometimes under CC-BY-NC. Evaluation and publication are fine; anything revenue-generating is not, until you buy a commercial licence — which for several labs is a real, listed product rather than a refusal. |
| Behavioural-use (RAIL family) | Permissive on the commercial axis, but with an appendix of prohibited uses that must flow to every downstream recipient. Compatible with a business; incompatible with the open-source definition, and awkward if you resell to customers you cannot bind. |
| Inherited licences on fine-tunes | A fine-tune generally cannot grant more than its base allowed. A community tune labelled Apache-2.0 sitting on top of a vendor-licensed base is a red flag, not a bargain — the uploader may simply have picked the wrong dropdown. Trace the chain to the original weights. |
Five questions before you ship
Run these against the actual licence file, in this order. The first one that fails is the one that would have cost you the project.
- May I use the output commercially? Separate from whether you may use the model. Some licences also constrain whether model output may be used to train other models — a clause that matters if you were planning to distil.
- May I redistribute the weights, and what must travel with them? If you ship an on-device app, you are redistributing. Notices, licence text and use policies usually have to go along.
- May I publish a derivative, and what must it be called? Several vendor licences impose naming and attribution rules on fine-tunes. Cheap to comply with, embarrassing to discover late.
- Is there a scale threshold? Some licences change character above a stated number of monthly active users. The thresholds are set high enough that they are irrelevant to almost everyone, which is precisely why nobody notices them until an acquirer’s lawyer does.
- What did the base model allow? Ask again at every level of the chain, including the dataset the tune was made with.
What this page is not
This is not legal advice, and no page can be. Licence texts are revised between model versions, the same family name can cover materially different terms across releases, and the consequences depend on your jurisdiction and your business. Read the LICENSE and any referenced use policy in the exact repository you are downloading from, keep a copy of both with the checkpoint hash, and take advice before you build revenue on a bespoke licence.
The habit worth forming is small: when you record a model in your stack, record its licence name and the URL you read it at in the same row. It takes seconds, and it is the difference between an answerable question and an archaeology project.