Open-weight models and the license file

"Open" is doing too many jobs in model announcements. Sometimes it means you can download the weights and run them on your own hardware, as in local models and API calls. Sometimes it means the license is permissive. Sometimes it means the paper was published and the weights were not. Those are different facts, and the download button only settles the first one.

This page is a map of the documents. It is not legal advice, and it is not a reading of any one vendor's license.

Two permissions people mix up

An API key lets you send prompts to a service under that service's terms. You do not have the weights. If the vendor withdraws the model id, the call stops. Their agreement covers retention, rate limits, and what you may build.

A weights file is a pile of numbers you can load in a runner. The permission to use it is the license that was distributed with the file. Having the file on disk, because a link worked, is not a substitute for that license. If the license and a blog post disagree, the license is the one that was offered with the artifact.

What the common shapes look like

Some models are released under permissive software licenses. MIT and Apache-2.0 are the names you will see. They allow broad use, and they still have conditions, including attribution and, for Apache-2.0, patent and notice terms. I am not restating those licenses here. Read the text that came with the weights.

Other models use a community license written for that release. These often allow research and many commercial uses, and then add restrictions. Typical restrictions include a cap tied to the size of the user base, a ban on using the weights to train a competing model, or a requirement to keep the same license on a derivative. The threshold, if there is one, is a number in that document. Do not take it from a recap.

A tag on a model hub can say "open" while the card links to a license with restrictions. Open the link. If there is no license file, you do not have a clear grant, and "it was on the internet" is not one.

What you usually do not receive

The weights are the result of training. The dataset is a separate collection with its own rights. Most downloads do not include that dataset. You cannot republish data you do not have, and you cannot assume the data was licensed for your use because the weights were.

The tokenizer and the inference code may be under a different license from the weights. Check each artifact if you are shipping them together.

A practical check before you build on a file

Find the license file that shipped with the exact revision you downloaded. Models are updated, and the next revision can change the terms. Match the commit or the version hash, not the product nickname.

Decide whether your use is the use the license describes. A prototype on one machine and a product for customers are easy to confuse in a README. If the license has a user cap or a commercial clause, this is the moment to have a lawyer read it, not the moment to trust a comparison table.

Keep the license text with the file in your own storage. A link to a repository can change. The copy you relied on should be the copy you can still open.

The Open Source Initiative publishes a definition of open source software. Use that definition when someone claims the phrase. Use the license file when you need to know whether your deployment is allowed. They answer different questions, and only the second one is about your build.

What to store next to the weights

When you download a revision, store four things in the same folder and do not rely on the repository still looking like that next month.

The weights file, named with the revision you ran. The tokenizer and any inference code you actually ship, each with the license that covered that artifact if it differs. The license text and the model card as they were on the day you downloaded them. A one-line note in your own words: prototype on one machine, or a product other people will use. The license cares about that difference even when the README treats them as the same experiment.

If a later download changes the license, the old folder is still the grant you accepted for the old file. Do not overwrite it and then argue from memory. The new file is a new grant, and it needs its own reading.

An API product beside this folder is a second pile: the contract, the model id, and the date you accepted the terms. Mixing the two piles is how a team says "we use the open model" while the production call is an API key. Write down which pile served the reply. The privacy question and the license question are both answered by that line.