Open Weight or Open Source? AI Model Licenses for Companies Explained

Strictly speaking, most "open-source" language models are open-weight models: The model weights are freely downloadable, but training data and training code are usually not. For companies, the license is crucial: Apache 2.0 – for example with gpt-oss, many Mistral models and the dense Qwen3 models – allows commercial use largely freely, while Meta's Llama license contains additional requirements.

If you use an open model in your company, you should know what you are allowed to do. This article explains the terms and the most important licenses. It does not replace legal advice, but it does provide guidance for the right question for your legal department.

Open Weight and Open Source – The Difference

• Open Weight: The trained weights – i.e. the finished model – are public. You can download it, operate it, and customize it often.
• Open source in the strict sense would also mean open training data and training procedures, so that the model could be replicated. This is rare for large language models.

In everyday life, however, one speaks of "open source models". In practice, what is more important than the term is what the license allows.

The most important licenses

The Llama license in detail

The Llama 3.3 Community License allows commercial use, but contains obligations:

• the model or products must display Anyone who passes on "Built with Llama" in a clearly visible way and enclose the license.
• If other models are trained and passed on with Llama outputs, their name must begin with "Llama".
• Meta's Acceptable Use Policy is considered part of the license.
• Companies with more than 700 million monthly active users need a separate license.

For a typical SME that uses Llama internally, the requirements are manageable. However, you should be aware of them, especially if you deliver a product to customers with Llama.

What companies should pay attention to

1. Check license per model version – it can change between versions.
2. Clarify the type of use: internal use, installation in a product, transfer to customers.
3. Consistently exclude non-commercial licenses.
4. Documenting which model is used for what in which version – this also helps with obligations under the EU AI Act.

It's also worth taking a look at the fine print when it comes to software: The article n8n license explains why n8n is not considered open source despite open code.

We check the license for your purpose with each model selection. An overview of the models we work with can be found on the Open Source Language Models page.

Frequently Asked Questions

What is an open-weight model?

An AI model whose trained weights are publicly available. It can be downloaded and operated by yourself. Training data and code, on the other hand, are usually not open.

Can I use Llama commercially?

Yes, the Llama license allows commercial use. In the event of distribution, conditions such as the notice "Built with Llama" apply; in addition, Meta's usage policy and a limit of 700 million monthly active users apply.

Which AI models are under Apache 2.0?

Among others, gpt-oss from OpenAI, many Mistral models such as Mistral Small 4 and the dense Qwen3 models. The license is valid for a specific version in each case.

Do I need a contract with the manufacturer for open-top models?

Not for use under Apache 2.0. If you use a model through a provider such as IONOS, its contractual terms also apply.

Would you like to know which model suits your project in terms of licensing law? Contact us via the contact page.