Small and on-device modelsMiddleLesson 2275 min read

Open-weight models and licences

Downloading a model does not mean you may use it any way you like. Learn the difference between open weights and open source, and how to read a model licence.

Lesson in motion

In 60 seconds

Open-weight models and licences

Downloading a model does not mean you may use it any way you like. Learn the difference between open weights and open source, and how to read a model licence.

1/4
In simple words
A friend lends you a cycle. You can ride it, but maybe not rent it out to others, and maybe not paint it. Downloading a model is like that: you get the model, but the licence says what you are allowed to do with it.
An open-weight model is one whose trained weights you can download and run yourself. That is different from a model you can only reach through someone's API. But "open weights" says nothing about what you may do with them. That is decided by the licence.

Open weights vs open source

TermWhat you usually getWhat it does not promise
Open weightsThe trained model files, so you can run and often fine-tune itThe training data, the full training code, or freedom to use it for anything
Open source (software sense)Code under a licence that allows use, study, change and sharing, for any purposeThat the model's training data is published
Open source AI (stricter sense)Weights, code and enough information about the data to rebuild or truly study the modelIt is still debated exactly where the line is
Many popular models are open-weight but not open source. They come with a custom licence that adds rules. The Open Source Initiative, the group behind the widely used open source definition for software, has also published a definition for open source AI. Not everyone agrees on it, so read each licence rather than trusting a label.

Five questions to ask every licence

  1. Commercial use: may I use it in a product that earns money? Are there limits, for example on company size or number of users?
  2. Attribution: must I name the model, keep a notice file, or show text like "built with …"?
  3. Acceptable-use rules: is there a list of banned uses? Do I have to pass these rules on to my users?
  4. Derivatives: if I fine-tune it, what licence covers my new model? Must I use a certain name?
  5. Outputs: may I use the model's outputs to train another model? Some licences and terms limit this.

Common licence families

  • Permissive licences such as Apache 2.0 and MIT: broad freedom, including commercial use, with conditions such as keeping the licence and notices.
  • Custom model licences: written by the model's publisher. They may allow commercial use but add acceptable-use policies, user limits or naming rules.
  • Non-commercial or research-only licences: fine for learning and experiments, not for a paid product.

Read the model card too

A model card is the documentation page for a model, for example on Hugging Face. It should say what the model is for, what data and languages it was trained on, known limits and risks, and which licence applies. The licence may be a separate file in the same repository.
Careless
  • "It's on the internet, so it's free to use"
  • Reading only the short licence name on the page
  • Assuming a fine-tuned copy has the same licence as the original
Careful
  • Opening and reading the full licence file
  • Checking the acceptable-use policy and any user or revenue limits
  • Checking both the base model's and the fine-tune's licences
Watch out
Licences and model pages change. Save a copy of the licence and note the date and the exact model version you downloaded. For a real business, ask a lawyer to review it. This lesson is general information, not legal advice.
Try this. Open any model page on Hugging Face. Find the licence name, then find the full licence text. How many clicks did it take?

Watch and read more

Lab

A licence checklist for two real open-weight models, and a choice for one project.

~30 min

The problem

Your project: "a paid app that helps small shop owners in Tamil Nadu write replies to customer messages in Tamil and English."

  1. On Hugging Face, pick two small open-weight text models. Open each model card and its full licence file.
  2. For each, fill in a table: licence name, commercial use allowed? (and any limits), attribution needed, acceptable-use policy (yes/no, and one example rule), rules about fine-tuned versions, rules about using outputs to train other models, languages listed on the model card.
  3. Write the date and the exact model name and version you checked.
  4. Choose one model for the project and give two reasons. Write one question you would ask a lawyer.

You are done when

Hard questions

Try to answer before you reveal. If you can answer these, you understood the lesson.

Q1A model's weights are under a permissive licence, but its model card says it was trained on data with unclear rights. Is it safe to use commercially?Reveal

Not automatically. The licence covers what the publisher grants you, but it cannot remove every risk that comes from the training data. For low-risk uses many teams accept this, but for a serious product you should read the card carefully, look for any legal notes from the publisher, and get legal advice. Keep a record of what you checked.

Q2Why might a licence ask you to pass its acceptable-use policy on to your own users?Reveal

Because the publisher cannot control what your users do with your app. Passing the rules down (for example through your terms of service) means the restrictions follow the model wherever it is used. If you ignore this, you may be breaking the licence even if you personally never misuse the model.

Please sign in to continue.

Questions people ask

If a model is on Hugging Face, can I use it commercially?

Not automatically. Hugging Face hosts models under many different licences. Some allow commercial use, some do not, and some add conditions. Always read the licence of that exact model.

What is an acceptable-use policy?

A list of things you may not use the model for, such as illegal activity or certain harmful uses. Some licences require you to make your own users follow the same rules, for example in your app's terms.

Does a fine-tuned model keep the original licence?

Usually the base model's licence still applies to your fine-tuned version, and it may say how you must name or share it. A fine-tune someone else published may also add its own terms. Check both.

Is using a model for a college project different?

Many licences are more relaxed about research and personal learning, and non-commercial licences are made for exactly this. But if the project later becomes a product or startup, check the licence again before you launch.

Lesson test

5 questions. Get 3 right (60%) to pass and complete this lesson.

Sign in to take the test and save your progress