The licence is a business decision

Three categories, and what each permits

Permissive: Apache 2.0, MIT. Use, modify, redistribute, sell, embed in a product. No conditions worth worrying about beyond keeping the notice. Many Qwen, Mistral and DeepSeek releases sit here. If you are building anything commercial, this is the easy category.

Custom community licences. Meta’s Llama licences and Google’s Gemma terms are the best-known examples. They permit commercial use, and they add conditions. Historically these have included an acceptable-use policy, a requirement to pass the terms downstream, naming requirements for derivative models, and (in Llama’s case) a clause requiring a separate agreement above a very large monthly-user threshold. For an individual or a small business these are almost never a practical obstacle. They are a real obstacle if you are a large platform, and they mean a lawyer should read them before you build a company on one.

Non-commercial and research-only. cc-by-nc, research licences, and evaluation-only terms. Less common than they were and still around, particularly for research releases and some specialist models. Read before you build.

Why “open source” is contested here

This is a genuine, unresolved disagreement, and it is worth understanding rather than picking a side.

The traditional definition of open source, maintained by the Open Source Initiative, requires freedom to use, study, modify and share without restrictions on field of use or on who may use it. Several widely-used model licences fail that test, usually because of acceptable-use restrictions or user thresholds, which is why the OSI and others argue those models should be called open weights rather than open source.

The counter-argument is that releasing weights that anyone can download, run and fine-tune delivers most of the practical freedom people care about, and that acceptable-use terms are a reasonable response to real risks.

There is a further dispute about whether a model can be open source at all when its training data is not published, since the data is arguably the source and the weights the build output. The OSI’s own definition for AI systems requires detailed data information rather than the data itself, which satisfies almost nobody completely.

None of this is settled. Different jurisdictions, standards bodies and companies use the terms differently, and the disagreement is ongoing. What matters for you is practical: read the actual licence rather than the word used to describe it.

The other unresolved question

Alongside the definitional dispute there is a legal one, and it is worth naming as unsettled rather than glossed. Models are trained on large amounts of text and images collected from the web, and whether that training is permitted without the rights-holders’ consent is being litigated and legislated differently in different countries. Some jurisdictions have text-and-data-mining exceptions; others do not; several cases are ongoing.

That uncertainty attaches to closed and open models alike, and nothing in a licence resolves it, because a licence governs what the maker permits you to do with the weights, not whether the weights were lawfully made. Anyone telling you this question is settled, in either direction, is describing their preference rather than the law. This is not legal advice; if the answer matters commercially where you are, it is a question for a lawyer in your jurisdiction.

The questions to answer before building

1. May I use this commercially? Stated plainly in permissive and non-commercial licences; requires reading in the custom ones.
2. Must I attribute, and how? Some require naming the base model in derivative names, or displaying “Built with X”.
3. Do the terms travel with it? If you fine-tune and release, are you obliged to pass the licence on?
4. Is there a threshold? Some terms change above a user count.
5. What does the acceptable-use policy prohibit? These read like usage policies for hosted services and they bind you as an operator.
6. What is the output’s status? Most licences do not claim ownership of what the model generates, but this is separate from whether that output is copyrightable at all, which differs by country and remains unresolved. This is not legal advice and the answer where you live may differ from the answer next door.

The practical advice

For learning, personal use, and internal tools, all three categories are usually fine and the distinction is mostly academic.

For anything you sell, embed in a product, or build a business on: prefer Apache 2.0 or MIT, because it removes an entire class of question. If a community-licensed model is better for your task, read the terms in full, keep a copy of the version you agreed to, and get advice if the numbers are large. The licences have changed between releases before, and the copy on the website today is not necessarily the one you accepted last year.