In the last lesson you downloaded a model with one command, no account, no payment. It’s easy to conclude that the model is simply yours now. For playing around at home, that’s close enough. The moment it goes into something at work, a tool for colleagues, a product, a customer project, the file comes with rules, and those rules are in a license that is easy to overlook.
This lesson is about reading that license. Not as a lawyer, but well enough to know whether a model is fine to use, or whether someone with a legal department should look first.
Three things a model release can contain#
A software project is “open source” when you get the source code, the thing you need to understand, change and rebuild it. A model has no single source. It has three parts:
- The weights: the billions of numbers that are the finished model. This is the file you download.
- The training code: the programs that prepared the data and trained the model.
- The training data: the text it learned from, or at least an exact description of it.
Most popular models publish only the first. You can run them and adjust them, but you can’t see how they were made, and you could never rebuild them yourself.
The Open Source Initiative, the body that has defined “open source” for software since the 1990s, published the Open Source AI Definition 1.0 in October 2024. It requires three things: the parameters (the weights), the complete code used to train and run the model, and enough information about the training data that a skilled person could build a substantially equivalent system. By that bar, most models called “open” are open weights, not open source. Ai2’s OLMo 3 is an example that goes all the way: weights, training data, training code and intermediate checkpoints are all public, under Apache 2.0.
For your daily work, this distinction matters less than the next one: what the license lets you do.
The license families you’ll meet#
Model licenses fall into a few groups. The first is the easy one.
- Permissive (Apache 2.0, MIT). The licenses that open source software has used for decades. Use it commercially, change it, ship it, as long as you keep the license and copyright notice. Apache 2.0 also wants you to note your changes and includes a patent grant. Qwen3, Gemma 4, OLMo 3 and gpt-oss use Apache 2.0, DeepSeek-R1 uses MIT.
- Custom “community” licenses. Written by the vendor, so each one is different. Meta’s Llama licenses are the best-known example: commercial use is allowed, but with conditions attached.
- Non-commercial. Research and personal use only. Cohere’s Command R models, for example, are under CC BY-NC 4.0, the “NC” being “non-commercial”.
Here are the models this site mentions, plus one non-commercial example:
| Model | License | Commercial use | What to watch |
|---|---|---|---|
| Qwen3 (all sizes) | Apache 2.0 | yes | keep the license notice |
| Qwen3.8-27B | Apache 2.0 | yes | the largest Qwen3.8 model has its own license |
| Gemma 4 | Apache 2.0 | yes | Gemma 1 to 3 used Google’s own terms |
| gpt-oss-20b | Apache 2.0 | yes | plus OpenAI’s gpt-oss usage policy |
| DeepSeek-R1 | MIT | yes | the distilled versions keep their base model’s license |
| OLMo 3 | Apache 2.0 | yes | data and code are open too |
| Llama 3.1 | Llama 3.1 Community License | yes, with conditions | user limit, use policy, naming rules |
| Nemotron 3 Nano | NVIDIA Open Model License | yes | a vendor license, read it |
| Command R | CC BY-NC 4.0 | no | plus Cohere’s acceptable use policy |
Checked against the model cards and vendor announcements in November 2026. Licenses change, so treat the table as a snapshot, not a promise.
What a community license looks like#
The Llama 3.1 license shows what to look for in any vendor license:
- A user threshold. A company whose products had more than 700 million monthly active users needs a separate license from Meta, which Meta may refuse. That affects only a few very large companies, but other licenses set the bar much lower. For EU readers: the use policies of Llama 3.2 and Llama 4 do not license their multimodal models to companies based in the EU; end users of a product built on them are not affected.
- An acceptable use policy. A separate document listing what you may not use the model for. Breaking it breaks the license.
- Attribution. If you distribute the model or a product built on it, you must show “Built with Llama” visibly.
- Naming rules. If you fine-tune Llama and publish the result, its name must begin with “Llama”.
None of that is unreasonable. But it’s a different kind of agreement from Apache 2.0, and someone in your organisation should have read it before you build on it.
Three pitfalls#
- The license is per model, not per family. Gemma 1 to 3 came with Google’s own terms; Gemma 4 switched to Apache 2.0 in April 2026. Qwen3.8-27B is Apache 2.0, the largest model of the same generation is not. Check the exact model you use.
- Derived models inherit. DeepSeek’s distilled R1 models are built on Qwen and Llama base models, and each keeps its base model’s license, not DeepSeek’s MIT. The same goes for any fine-tune you find on the internet.
- Repackaged files carry the original license. A GGUF file uploaded by a third party is still bound by the license of the model it was made from, whatever the third party’s repository says.
Five questions before a model goes to work#
- Is commercial use allowed? If the answer is “no” or “research only”, stop here.
- Is there a size limit? Monthly users, revenue, or anything else that might apply to you later.
- Is there a use policy? Read it once and check that your use case isn’t on it.
- What do you have to show? License text, copyright notice, “Built with …” badge.
- What if you change it? Fine-tunes may need a name prefix or the same license.
Usually fine without further thought
- Apache 2.0 or MIT, used internally
- Trying a model at home or on a test machine
- Running it unchanged behind your own tool
Ask someone who reads contracts
- Any custom license in a product for customers
- Publishing your own fine-tune
- Anything marked non-commercial, outside of research
Where to find the license#
On Hugging Face, the license sits in the tag row at the top of the model page, as in
License: apache-2.0. Under “Files and versions” you’ll usually find a LICENSE file with
the full text. Models with a custom license are often gated: you have to accept the terms
in a form before the download starts, which is a good moment to read them.
If you downloaded through Ollama, the terminal shows it too:
ollama show qwen3:8b --license
This lesson isn’t legal advice, and I’m not a lawyer. But the five questions above catch nearly every case that matters, and they take five minutes. Spend them before the model goes into something other people depend on, not after.