Can you use Llama, Mistral or Gemma in a commercial product?

Open-weight models are language models whose trained parameters, the "weights", you can download and run on your own servers. That freedom is why many small companies look at Meta's Llama 3, Mistral AI's open models and Google's Gemma when they want a chatbot or writing tool they control. But downloading a model is not the same as owning it. Each comes with a licence, a legal text that says what you may do with it, and the three licences are quite different in tone and detail.

This article summarizes what each text says as of early May 2024, then walks through a sample check. It is a reading guide, not legal advice: licences change, your situation matters, and a lawyer who has read the full text is the person to confirm your plans.

Mistral's open models: Apache 2.0

Mistral AI released Mistral 7B and Mixtral 8x7B under the Apache License 2.0, a general-purpose open-source licence published by the Apache Software Foundation in 2004. Its main points:

You receive a permanent, worldwide, royalty-free right to copy, change, distribute and sublicense the work, including in paid products.
Contributors also grant you a licence to their relevant patents; that grant ends if you sue claiming the work infringes a patent.
If you pass the model on, you include a copy of the licence, keep the copyright and attribution notices, flag files you changed and keep any NOTICE file.
It gives no right to use the licensor's trademarks, beyond saying where the work came from.
The work comes "as is", with no warranty.

There is no cap on users and no list of banned uses inside the licence. Note that this applies to Mistral's open-weight releases. Its flagship Mistral Large is offered through Mistral's own platform and Microsoft Azure, under separate terms.

Llama 3: the Meta Llama 3 Community License

Meta's licence, dated April 18, 2024, grants a worldwide, royalty-free right to use, copy, modify and distribute Llama 3, with conditions attached:

Attribution. If you distribute Llama 3 or offer a product that uses it, you must include the licence and show "Built with Meta Llama 3" prominently, for example on your website or in your product documentation. Copies must include a Notice file with Meta's credit line.
Naming. If you fine-tune or otherwise build a model from Llama 3 and release it, its name must start with "Llama 3".
Acceptable Use Policy. Your use must follow Meta's policy, which forbids, among other things, illegal activity, harassment, unlicensed professional practice, impersonation and presenting outputs as human-made.
No training rivals. You may not use Llama 3 or its outputs to improve other large language models.
Very large companies. If your products counted over 700 million monthly active users in the calendar month before the release date, you need a separate licence from Meta.

Gemma: the Gemma Terms of Use

Google's terms, last modified February 21, 2024, allow use, modification and distribution, also with conditions:

Distribution is defined broadly. Offering Gemma through an API or website counts, not only handing over the files.
Pass on the restrictions. Anyone you distribute Gemma or a modified version to must be bound by Google's use restrictions, receive a copy of the terms and see which files you modified. Non-hosted copies need a Notice file.
Prohibited Use Policy. You may not use Gemma for the uses listed in Google's separate policy, and Google may restrict use it reasonably believes breaks the terms.
Outputs are yours to manage. Google claims no rights in what Gemma generates; you and your users are responsible for it.
Keep current. You are expected to make a reasonable effort to stay on Google's newest Gemma release.

A worked compliance check

Imagine a small accounting firm (made up for this example) that wants a chatbot on its website to answer common questions about its services, running on its own server.

Question	Llama 3	Mistral 7B	Gemma 7B
Commercial use allowed?	Yes	Yes	Yes
Attribution needed on the site?	Yes: "Built with Meta Llama 3"	Licence and notices if distributed	Terms and Notice file if distributed; restrictions passed on
Use policy to follow?	Meta's Acceptable Use Policy	None in the licence	Google's Prohibited Use Policy
Does the chatbot give tax advice?	Check the policy on unlicensed professional practice	Licence silent; other laws still apply	Check the prohibited uses
User cap?	700 million monthly users	None	None
Fine-tuned model naming?	Must start with "Llama 3"	No rule	No rule

For this firm, every model is possible. The practical steps are to save a dated copy of the licence and policy it relies on, add any attribution, have the chatbot avoid giving personal tax advice, and review the terms when it upgrades the model.

Try it: Pick one model you are considering and open its full licence. Make a five-line list: what you may do, what you must show, what you must not do, any user limits, and what happens if you break the terms. Bring that list to your lawyer or adviser.

How do their licences differ?

The biggest difference is philosophy. Apache 2.0 is a general open-source licence: it grants wide rights, asks for notices and attribution when you redistribute, and says nothing about how you use the software. The Llama 3 and Gemma terms were written for AI models specifically, so they add use policies that list forbidden purposes and reserve the right to act on violations.

The second difference is scale and competition. Only Llama 3 has a user threshold, 700 million monthly users, which matters to a handful of very large companies. Only Llama 3 forbids using the model or its outputs to improve other language models. Gemma's definition of derivatives reaches models trained to imitate Gemma, such as through distillation, which brings them under its terms.

The third is visibility. Llama 3 requires a public "Built with Meta Llama 3" credit and a naming rule for derived models; Gemma and Apache require notices that travel with the files. In every case, the licence text you accepted is what counts, so keep a copy, and ask qualified counsel when the stakes are real.