Llama, Mistral or Qwen? Choosing open-weight models for your organisation
Short answer
For business use in Europe the licence weighs as heavily as the quality. Mistral (French) and Qwen (Alibaba) release their recent open-weight models under Apache 2.0, without usage restrictions. Meta’s Llama has its own licence, which grants no rights to the multimodal models, including Llama 4, to companies with their principal place of business in the EU. Then choose by language, task and the hardware you have.
Licence first: what may you do with it?
| Model family | Maker | Licence (recent models) | Watch out |
|---|---|---|---|
| Mistral | Mistral AI, Paris | Apache 2.0 (Mistral 3 family, Mistral Small 3.x) | Older or specialised models may have a different licence |
| Qwen | Alibaba, China | Apache 2.0 (Qwen3) | Some organisations weigh the origin in their supplier policy |
| Llama | Meta, US | Llama Community License with acceptable use policy | No rights to multimodal models for EU companies; separate licence above 700 million monthly users |
Llama’s EU restriction deserves attention. Since Llama 3.2, the acceptable use policy states that the rights to the multimodal models are not granted to individuals domiciled in the EU or companies with their principal place of business there. End users of a product that incorporates such a model are exempt, but a European organisation that wants to run the model itself is covered. Llama 4 is multimodal. The text-only models, such as Llama 3.3 70B, are not covered by this restriction.
Apache 2.0 is a well-known, permissive licence: commercial use, modification and redistribution are allowed, with attribution. For a legal department it is the simplest starting point.
Then: language, task and size
- Language. Test in the languages you work in, with your own texts. Quality in Dutch, German and French differs by model and version.
- Task. Small models (a few billion parameters) are fast and cheap for classifying, summarising and extracting. Large models write and reason better. Many organisations run more than one.
- Size and hardware. A 30-billion-parameter model runs in 4-bit on one datacentre GPU; a 70-billion model too, but with little headroom. Large mixture-of-experts models, such as Mistral Large 3 with hundreds of billions of parameters, need several GPUs. See An on-premises LLM server.
- Context length. If you want long contracts or files processed in one go, look at the context window, and what it costs in memory.
Does a model’s origin matter?
An open-weight model running on your own hardware sends no data to its maker. It is a file of weights, not a service. The data risk therefore does not depend on where the maker is based.
Origin can matter for other questions: which values and taboos the model picked up in training, whether your organisation’s or your client’s policy excludes suppliers from certain countries, and how transparent the maker is about training and safety. Record the trade-off, so you can explain it to an auditor or client. Under the AI Act, transparent model provenance also helps your own documentation. See The EU AI Act in 2026.
Do not choose forever: keep models interchangeable
Better models come out every six months. The key design principle is therefore that you can switch models without rebuilding your automations. That works if your applications talk to a standard API, and you can decide per role which model sits behind it.
In werqly AI Control you choose per role which model is available: a fast small model for the service desk, a larger one for legal analysis. werqly tests new open-weight versions and rolls them out to your node; you decide when a role moves over. The models werqly supplies are chosen with the licence terms for European organisations in mind. More on choosing your own environment in What is private AI?
Frequently asked questions
Questions on this topic
May a European company use Llama 4?
According to Meta’s acceptable use policy, the rights to the multimodal Llama models, including Llama 4, are not granted to companies with their principal place of business in the EU. End users of a product that incorporates the model are exempt. Have the terms checked legally for your situation.
What does Apache 2.0 mean for business use?
Apache 2.0 allows commercial use, modification and redistribution, with attribution and without usage restrictions such as a maximum number of users. Mistral’s and Qwen’s recent open-weight models are released under this licence.
Is a Chinese open-weight model a data risk?
Not when it runs on your own hardware: the model is a file and sends no data to its maker. Origin can play a role in your supplier policy and in questions about the model’s training and values. Record that trade-off.
Which open-weight model is best?
That depends on language, task, licence and hardware, and changes every six months. Test two or three candidates with your own documents and tasks, and set up your environment so you can switch later without rebuilding automations.
Sources
- Meta — Llama 4 Acceptable Use Policy (GitHub)
- Mistral AI — Introducing Mistral 3 (December 2025)
- Qwen — Qwen3 on GitHub
- Sara Zan — Using Llama models in the EU (2025)
This article is general information, not legal advice. Rules and guidance change; check the sources listed or ask your lawyer or DPO if in doubt.