Can I use an open-weight model like Llama or Gemma in my product?
What to check before building on an open-weight model, including license type, Llama's user cap, attribution and naming rules, and EU limit, acceptable use policies, fine-tuning, and the protections you give up compared with a hosted API.
Short answer
Usually yes, but "open" doesn't mean the same thing for every model. Some open-weight models are released under standard permissive licenses like Apache 2.0 or MIT, which allow commercial use with few conditions. Others use custom licenses with their own rules, such as user caps, attribution requirements, naming rules, geographic limits, and acceptable use policies that carry over to your users. Read the license for the exact model and version you ship, because terms can change between versions of the same family. Checking this before you build lets you pick the model that fits your business, instead of discovering a restriction at launch or in diligence.
Two kinds of licenses
- Standard permissive licenses. Apache 2.0 and MIT allow commercial use, modification, and redistribution. The main conditions are keeping copyright and license notices, and, for Apache 2.0, noting changes to files you modify. Apache 2.0 also includes a patent license. Google's Gemma 4 models, released in April 2026, use Apache 2.0, a change from the custom license on earlier Gemma versions.
- Custom model licenses. Meta's Llama licenses are the best-known example. They allow commercial use but add conditions you have to follow, and the license on one version may differ from the next.
What custom licenses can require
Using the Llama 4 Community License as an example:
- A user cap. If your products and affiliates had more than 700 million monthly active users on the model's release date, you need a separate license from Meta, which it may grant or refuse. Most companies are far below this, but it matters for acquirers and large partners.
- Attribution. If you distribute the model or offer a product that uses it, you must display "Built with Llama" on a related website, user interface, blog post, about page, or product documentation, and include the license's notice with copies you distribute.
- Naming. If you use Llama to create, train, or fine-tune a model you distribute, its name must begin with "Llama."
- An acceptable use policy. Meta's policy prohibits certain uses, and your users' use of your product has to stay within it. Your terms of service should pass those limits on. See What should the terms of service for an AI product cover?
- A geographic limit. The Llama 4 acceptable use policy doesn't grant rights to the multimodal Llama 4 models to individuals living in the EU or companies with their principal place of business there.
Other custom licenses have their own variations, such as revenue thresholds, separate agreements for large hosting companies, or non-commercial terms. Treat each model's license as its own contract.
Other things to check
- What you give up compared with an API. Open models come as-is, without warranties. Some hosted AI providers offer contractual protections, such as copyright indemnities for certain enterprise plans, that you won't get from downloaded weights.
- Training data questions. A model's license grants rights from the developer, but it doesn't settle disputes over the data the model was trained on. Courts are still working through those cases. See Who owns AI-generated output?
- Fine-tuning and the EU. Under the EU AI Act, a company that substantially modifies a general-purpose AI model and offers it in the EU can take on the original provider's obligations for that modification. Under the European Commission's guidance, that generally happens when the fine-tuning uses more than a third of the compute used to train the original model, a level most product fine-tuning doesn't reach.
- Customer requirements. Enterprise and government customers may ask which models you use and where they came from, and some restrict models from particular developers or countries. Know your answer before the security questionnaire arrives.
- Your own privacy promises. Running a model on your own servers can keep user data away from a third-party provider, which can simplify your privacy story. If you rely on a hosting provider, its terms apply too. See Can I send user data to an LLM provider?
- Security of model files. Some model file formats can run code when loaded. Download weights from the official source, prefer safe formats such as safetensors, and record versions and checksums.
Checklist an agent can run against the codebase
- List every open-weight model the product uses, including embedding, speech, and vision models, with the exact version and where it was downloaded from.
- Find the license for each model and version, and note whether it's a standard license or a custom one.
- For custom licenses, list the conditions: user or revenue caps, attribution, naming, geographic limits, and acceptable use policies.
- Check that required attribution, such as "Built with Llama," appears in the product, website, or documentation.
- Check that license and notice files ship with any distributed weights or fine-tuned models, and that fine-tuned model names follow the license.
- Check that the terms of service pass on the acceptable use policies of the models in use.
- Check for EU availability of any model whose license excludes EU users or companies.
- Check model file formats, and that downloads are pinned to a version and verified with checksums.
Sources
- Llama 4 Community License Agreement
- Llama 4 Acceptable Use Policy
- Gemma 4 model card on Hugging Face (Apache 2.0)
- Apache License 2.0
If this applies to you
An AI feature review covers the models behind a feature before launch: each license's conditions, the acceptable use terms to pass on to users, attribution, and how the model choice fits your privacy documents and customer contracts. For one decision, such as switching to a new open-weight model, a pre-build feature check covers that choice before you build on it.