Open weights is not open source

Two phrases that are not the same

Open weights means the numbers are published. You can download the file. That is the whole of it.

Open source is not a mood. It is a written definition, kept by the Open Source Initiative. A licence either meets it or it does not.

Two of its rules matter most here. A licence must not discriminate against any person or group. And it must not stop anyone using the work in a particular field of endeavour.

So a file anyone can download may still carry rules about who you are and what you do with it. The download tells you nothing. The licence tells you everything.

Backed by The Open Source Definition.

The plain ones: Apache-2.0 and MIT

Some makers use the same licences that ordinary software uses. Qwen ships the Apache License 2.0 text with Qwen3-8B. Microsoft ships the MIT License with phi-4. DeepSeek-R1 is MIT. Ai2's OLMo 2 and OpenAI's gpt-oss are Apache-2.0.

MIT is short enough to read in a minute. It lets you use, copy, change, merge, publish, sell and pass on the work without restriction. It asks one thing in return: keep the copyright line and the licence text with any copy.

Apache-2.0 gives the same freedom and adds a patent licence. It asks a little more when you pass the work on. Give people a copy of the licence. Mark the files you changed. Keep the notices that are already there. Pass on the NOTICE file if one exists.

Neither licence grants you the maker's trademarks. And neither asks what field you work in, how large your company is, or what you plan to build. That last part is the difference that matters.

Backed by The MIT License, Apache License, Version 2.0, Qwen3-8B LICENSE file, Phi-4 LICENSE file, DeepSeek-R1 model card, OLMo 2 7B model card, gpt-oss-20b model card.

Llama's licence has conditions

Meta's Llama models come with Meta's own licence. It is generous, and it is not Apache-2.0. Here is what the Llama 3.3 version asks of you.

• Pass on a copy of the agreement if you share the model, or a product that contains it.
• Display "Built with Llama" prominently on a related website, interface, blog post, about page or documentation.
• If you train another AI model on Llama or its outputs, and then release that model, its name must begin with "Llama".
• Keep a Notice file carrying Meta's attribution line in any copy you distribute.
• Follow the Acceptable Use Policy. The licence pulls it in and makes it part of the agreement.
• If your products had more than 700 million monthly users in the month before the release date, ask Meta for a licence first.

Backed by Llama 3.3 Community License Agreement.

When the maker writes its own rules

Llama's Acceptable Use Policy is a separate document, and the licence makes it binding. Some of it closes off whole fields. Military, warfare and nuclear uses. Espionage. Guns and weapon development. Illegal drugs and controlled substances.

One part is narrower than it first looks, and the difference matters. It bans the unauthorised or unlicensed practice of a profession, and it names financial, legal and medical work. That is a rule about practising without a licence. It is not a ban on the subject.

There is a way to read a policy like this that saves time. Do not ask whether your work sounds risky. Ask whether the list names your field. It is specific, and most work is not on it.

One phrase widens it, though. The policy says you will not use, or allow others to use, Llama for the things it lists. So the rules follow anything you build and hand on.

Backed by Llama 3.3 Acceptable Use Policy.

Terms that travel with the file

Google's Gemma terms travel with the file as well. Pass on a copy of the agreement. Mark any files you changed. Carry the use restrictions forward into your own agreement, as a term you can enforce.

Google adds one thing in your favour. It says it claims no rights in the outputs you generate with Gemma. It also keeps one thing for itself, and that one is worth reading twice.

Those Gemma terms now cover only the models listed in their own appendix. Gemma 4 has a separate licence page, and that page is the Apache License 2.0. A maker can step back as well as forward.

Then there is research only. Mistral AI publishes some models under Apache-2.0 and others under the Mistral AI Research License. That one limits use to non-profit, non-commercial research. It says outright that an employee using it in daily work does not count.

Same maker. Same download page. Different permission. Check the model, not the brand.

Backed by Gemma Terms of Use, Gemma 4 license, Mistral-7B-Instruct-v0.3 model card, Mistral AI Research License.

One rule that is about where you live

The Llama Acceptable Use Policy ends its list of rules with a sentence that has nothing to do with what you build. It is about where you are.

It covers any multimodal models included in Llama 3.3. For those, the licence's main grant is not given to an individual domiciled in the European Union. A company whose principal place of business is there is treated the same way.

There is a limit on that limit. Somebody who merely uses a product built on such a model is not caught by it.

For Llama 3.3 the sentence bites nothing. Meta's own model card describes Llama 3.3 as text only, so there are no multimodal models in it to restrict.

It is not hypothetical, though. The same sentence sits in the Llama 3.2 policy. Llama 3.2-Vision is a real collection, at 11B and 90B, that takes images and text in.

Now hold that against the rule at the top of this lesson. An open source licence must not discriminate against any person or group of persons. This one turns on where a person lives.

Backed by Llama 3.3 Acceptable Use Policy, Llama 3.3 model card: a text only 70B model, Llama 3.2 Acceptable Use Policy, Llama 3.2-Vision model card: multimodal 11B and 90B models, The Open Source Definition.

What happens if you sue

Not every licence says what happens if you sue. The ones that do are short, they go unread, and they are not the same.

MIT does not mention patents at all. It grants none and takes none away. That gap is one reason Apache-2.0 exists.

Apache-2.0 does grant a patent licence. Start patent litigation claiming the work infringes, and your patent licences for that work end the day the case is filed.

Notice what survives. Apache ends the patent part only. The right to use, change and pass on the work is not what that clause takes back.

Meta's clause is wider. It covers a claim against Meta or anyone, about the Llama materials or their outputs. File one, and every licence the agreement gave you ends that day.

Almost nobody reading this will ever file such a claim. It is here because plain does not mean the licence says nothing.

Backed by The MIT License, Apache License, Version 2.0, Llama 3.3 Community License Agreement.

What is still being argued

Some of this is settled fact. Some of it is not, and it is worth being clear about which is which.

The Open Source Initiative says Meta's Llama licence is not open source. It names two reasons: the limit on commercial use for some users, and the Acceptable Use Policy. That post was written about Llama 2, and both of those conditions are still in the Llama 3.3 licence.

You will see these models called open source in plenty of places anyway. Check the licence rather than the word.

The Open Source Initiative has also published a definition for AI. It asks for three things. Enough detail about the training data for a skilled person to build something equivalent. The code used to train and run the system. And the parameters themselves.

Very few models clear that bar. Ai2's OLMo publishes code, checkpoints and training details, which gets it closer than most.

One thing here is genuinely unsettled. That same definition says it does not demand any particular legal mechanism to keep model parameters free. It expects this to get clearer once the legal system has had more chance to deal with these systems. The law is still being worked out.

Backed by Meta's LLaMa license is not Open Source, The Open Source Definition, The Open Source AI Definition 1.0, OLMo 2 7B model card, Llama 3.3 Community License Agreement.