Gemma's Licence: What Google's Terms of Use Actually Restrict

Gemma weights are free to download and free to use commercially. They are not open source, and the difference is not pedantry: the licence carries obligations that follow the weights downstream, and one of them is a document that Google can update without asking you.

It is not an open-source licence

Gemma is released under the Gemma Terms of Use, a bespoke Google agreement published at ai.google.dev/gemma/terms. It is not Apache 2.0, not MIT, and it does not meet the Open Source Definition, because it restricts what you may use the software for. An OSI-approved licence may not do that.

This matters practically rather than philosophically. Corporate policies, procurement questionnaires and dependency scanners often treat “open weights” and “open source” as the same category, and Gemma will fail an audit that assumes it. If you need a genuinely permissive licence for an open-weight model, families released under Apache 2.0 are the comparison to make, on licensing grounds alone.

What the grant gives you

The terms grant a non-exclusive, worldwide, royalty-free right to use, reproduce, modify and distribute the model and derivatives of it, including for commercial purposes and including hosting it as a service. There is no revenue threshold and no monthly-active-user trigger of the kind some other open-weight licences carry. There is also no field-of-use carve-out for research versus production.

So the everyday questions have easy answers. You may fine-tune it. You may serve it to paying customers. You may put it in a product you sell, and on a device you ship. You do not owe Google a fee or a report.

The warranty and liability position is what you would expect and worth knowing anyway: the model is provided as-is, with no warranty of any kind, and Google disclaims liability for what it produces or for what you do with it. That is normal for open weights, and it is the reason the obligation to evaluate the model for your own use case sits entirely with you. It also means the safety tuning in the instruction-tuned checkpoints is a property Google describes, not a guarantee Google underwrites.

The use restrictions, and where they live

The restrictions are not written out in the agreement. The agreement incorporates a separate document, the Gemma Prohibited Use Policy, published at ai.google.dev/gemma/prohibited_use_policy, and binds you to whatever it says. That indirection is the single most consequential structural feature of the licence and it is easy to read past.

The policy prohibits the familiar categories: generating or facilitating child sexual abuse material, non-consensual intimate imagery, harassment and abuse, malware and other harmful code, content that facilitates violence or the development of weapons, coordinated inauthentic behaviour and disinformation, circumvention of safety filters, and deceptive impersonation. It also reaches into professional-advice territory, restricting unqualified medical, legal and financial advice presented as authoritative.

Two features of the restrictions do real work in a compliance review. Google reserves the right to update the policy, and because the agreement references it rather than quoting it, an update changes your obligations for weights you downloaded a year ago. And Google reserves the right to require you to stop using or distributing the model if it reasonably believes it is being used in violation. That is a remote off-switch in contract form, not in code, but it is there.

How binding any of that is in practice is a question about enforcement, and the honest answer is that the mechanism is contractual rather than technical. Nothing in the weights checks a policy, and a downloaded checkpoint keeps working whatever Google later decides. What the terms give Google is standing: a documented agreement that a user accepted, with a defined breach and a defined remedy. For a business that is the part that matters, because the risk is not that your inference stops one morning. It is that a customer, an acquirer or a regulator asks which licence your model is under, and the answer turns out to have a policy attached that nobody on the team has read.

What you owe when you redistribute

If you pass the model on, whether as raw weights, a quantised repackaging, or a fine-tune, three obligations attach.

• Pass on the terms. Recipients must be bound by the same restrictions. You cannot relicense a Gemma derivative under Apache 2.0, and a fine-tune published without the terms attached is a licence breach by the publisher, not by the person who downloads it.
• Include the notice. Distributions must carry a notice stating that the material is provided under and subject to the Gemma Terms of Use, with the location of those terms.
• Mark your modifications. Modified files must carry a prominent notice that you changed them, in the ordinary style of such clauses.

The obligation that most often surprises people is the first one, because it is not what a permissive licence trains you to expect. A Gemma fine-tune is a Gemma derivative and inherits the use restrictions permanently. There is no laundering step.

Note what does not count as distribution. Serving the model over an API you operate is use, not redistribution: your users receive outputs, not weights, so the notice obligations do not reach them. Shipping the weights inside a desktop application, a mobile app bundle or a container image you hand to a customer is redistribution, and it does. The line is whether the recipient ends up in possession of the model, and it is the question to ask about any on-device or self-hosted deployment story.

Outputs, derivatives and fine-tunes

Google does not claim ownership of what the model generates for you. The terms are explicit that outputs belong to whoever generated them, subject to the same use restrictions, which is a materially friendlier position than some hosted-model terms take.

What counts as a derivative is broader than a fine-tuned checkpoint. Quantisations, merges, distillations that use Gemma outputs as training data, and adapters distributed together with base weights all sit inside the definition in practice. If you are building a pipeline that trains a new model on Gemma-generated data and expects to release it under a permissive licence, that is the clause to have a lawyer read.

The synthetic-data case deserves the emphasis because it is where teams most often assume the restrictions stop. If your training corpus was generated by Gemma, the resulting model is not obviously independent of Gemma’s terms, and a merged model that includes any Gemma-derived weights carries them for certain. Practically, that means recording the provenance of every training set and every merge input while you still remember it, rather than reconstructing it under a due-diligence deadline two years later.

What to check before you ship

1. Read the current text at ai.google.dev/gemma/terms and the current prohibited-use policy, and archive both with the date. Your obligations are the versions in force, not the ones in a blog post.
2. Confirm your use case is not in a restricted category, and write down the reasoning. “We generate medical summaries” needs a paragraph, not a shrug.
3. If you redistribute anything Gemma-derived, attach the terms, the notice and a modification statement to the artefact itself, not just to a README on a website.
4. Record the exact revision you shipped, so that a later licence question can be answered about a specific artefact. The pinning walkthrough covers the mechanics.