Licensing: What You May Legally Fine-Tune and Deploy
6 min read · updated August 3, 2026
This page is not legal advice, and nothing in it is a substitute for reading the licence attached to the specific checkpoint you are using or for asking a qualified lawyer. Licence texts are revised, differ between releases in the same family, and sometimes differ between sizes in the same release. What follows is a map of what to read and which clauses tend to matter.
Four separate licences, not one
“Can I use this model commercially?” is four questions wearing one coat. Each layer has its own terms and each can independently block you.
| Layer | Description |
|---|---|
| The base weights | The licence on the checkpoint you start from. This is the one people check. |
| The training data | The licence on the dataset you fine-tune with. A public instruction dataset can carry a non-commercial licence even when the model does not. |
| The teacher's outputs | If any training data was generated by another model, that provider's terms of use apply to those outputs — a contract you accepted, separate from any weight licence. |
| Your distribution | What you are permitted to do when you ship: hosting it as a service, redistributing weights, redistributing an adapter, or shipping it inside a product on a device. |
The most common expensive mistake in this area is a permissively licensed base model fine-tuned on a dataset that was generated by a commercial API. The weights were fine. The data was not.
What the licence families require
Read this as a taxonomy of what to look for, then read the actual licence file in the repository of the checkpoint you are using.
| Family | Description |
|---|---|
| Apache-2.0 / MIT | Genuine open-source. Commercial use, modification, redistribution and relicensing of your derivative are all permitted, with no acceptable-use restrictions and no user thresholds. Apache-2.0 additionally grants patent rights and requires you to preserve attribution and NOTICE files. Mistral's 7B and Mixtral 8x7B releases and DeepSeek's R1 weights are examples of models released under these terms. |
| Meta Llama Community Licence | Not OSI open-source, and each Llama release has its own version of the text. Broadly permissive commercially, with specific obligations: display attribution, include the licence when you redistribute, comply with and pass on an Acceptable Use Policy, and — above a very large monthly-active-user threshold (700 million in the Llama 2 and Llama 3.x texts) — obtain a separate licence from Meta. There is also a derivative-naming requirement, covered below. Some releases have carried additional territorial restrictions on particular modalities. |
| Google Gemma Terms of Use | Also not OSI open-source. Commercial use and derivatives are permitted, but you must pass the Terms and the Prohibited Use Policy on to anyone you distribute the model or a derivative to, and mark modified files. The use restrictions travel with every derivative, so you cannot licence your fine-tune more permissively than the base. |
| Bespoke vendor licences | Several families ship a custom licence, sometimes with a user or revenue threshold, sometimes restricting hosted-service use, sometimes research-only. The trap is that these vary by size and by release within one family — one model in a family can be Apache-2.0 while another is under bespoke terms. Check per checkpoint, not per vendor. |
| Non-commercial (e.g. CC BY-NC) | Research and evaluation only. No commercial deployment of the model or anything derived from it, including an adapter. Appears on research releases and, importantly, on many public instruction datasets. |
Naming, notice and passing terms along
Three obligations show up repeatedly in the non-OSI licences and are the ones most often missed, because they apply at the moment you ship rather than at the moment you train.
- Derivative naming. The Llama community licences require that a model you distribute which was trained on Llama materials includes “Llama” at the beginning of its name. If you are shipping a fine-tune as a product artefact, the product name is affected.
- Attribution notice. The same licences require a visible “Built with Llama” style attribution on associated materials. This is a documentation and UI task, not just a file in a repository.
- Passing terms downstream. Both the Llama and Gemma licences require that the terms and the acceptable-use policy travel with any redistribution. You cannot strip them, and you cannot offer your derivative under more permissive terms than you received.
A specific question worth resolving early: does an adapter count as a derivative? A LoRA adapter is meaningless without the base weights, and licences generally speak of models trained on or derived from the licensed materials. Treat the adapter as carrying the base model’s obligations unless you have advice saying otherwise — that is the conservative reading and it costs you almost nothing.
Training on another model’s outputs
This is contract law rather than licensing, and it is the layer most often skipped. When you call a commercial API you accept terms of use, and those terms have historically included restrictions on using the outputs to develop or train models that compete with the provider. Whether a particular training use falls inside such a restriction depends on the exact wording, which vendor, and which date — the terms are revised.
Three practical notes. First, read the current terms of the specific provider before generating a corpus, not after. Second, the restriction travels with the data: a public dataset generated from a commercial model carries the consequence to you, which is why Stanford released the Alpaca dataset under a non-commercial licence. Third, the direction has not been uniform — some open-weight licences have moved the other way, with Meta’s Llama 3.1 licence permitting use of outputs to improve other models subject to the naming requirement, where Llama 2’s text had restricted it. Do not assume the position you learned last year still holds.
The dataset licence is separate
Public instruction datasets are the quiet risk, because they are easy to download and their licences are not shown next to the weights.
- Check the licence on every dataset, including each component of a mixture. Aggregated datasets frequently combine sources under different terms, and the aggregate is bound by the most restrictive one.
- Datasets generated from commercial models are common and often carry non-commercial licences for that reason. Alpaca is the best-known example and far from the only one.
- Your own production data has its own constraints — your privacy policy, your customer contracts, data-protection law, and any data-processing agreements. “We may use your data to improve our services” is not obviously the same permission as training a model on it, and for regulated data it is usually not.
- Record provenance per row. A dataset whose origins you cannot reconstruct is one you cannot clear later, and later is when someone will ask.
A checklist before you ship
- Read the licence file in the base checkpoint’s own repository — the specific size, the specific release. Not a summary, not a blog post, not this page.
- List every dataset, including generated ones, with its licence and its origin.
- If any training data came from a commercial API, read that provider’s current terms on model training.
- Check for user or revenue thresholds and confirm you are below them — and note what happens if you cross one.
- Confirm your deployment mode is covered: hosting as a service, redistributing weights and shipping on-device are frequently treated differently.
- Implement the naming, attribution and pass-through obligations in the product, not just in the repository.
- Have a lawyer review it. Everything above is a map of what to bring them, not a substitute for the conversation.