Open-Weight Licences Compared: What Can Your Business Do?

Downloading an open model costs nothing. What you may do with it is set by a licence that differs per model, and since 2025 those licences increasingly carry conditions that hit a European company of 50 to 500 people directly.

Open-weight licences are the terms under which a maker releases the weights of an AI model: they decide whether you may use the model commercially, modify it, build it into a product or offer it to others as a service. Some are standard open-source licences such as Apache 2.0 or MIT. A growing share are the maker's own licences, with revenue or user thresholds, mandatory attribution and, in Llama 4's case, an exclusion of companies in the European Union.

Which open-weight licences apply in October 2026?

The open-weight licences of the major models fall into three groups in October 2026: free licences without conditions, custom licences with thresholds most companies don't reach, and licences with a condition that can catch you.

Checked on 17 October 2026 against the licence texts on Hugging Face and the makers' sites. Licences change per version: the same model name with a new version number can carry a different licence.

Can a European company use Llama 4?

In practice, a company based in the EU can't use Llama 4 itself. The Llama 4 Acceptable Use Policy, which is part of the licence, says: "With respect to any multimodal models included in Llama 4, the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union."

That reads like an exception for the image models only. But Meta's model cards describe the whole Llama 4 family as "natively multimodal AI models". Scout and Maverick are both covered. The carve-out is narrow: the restriction "does not apply to end users of a product or service that incorporates any such multimodal models". A Dutch customer may use a US service running on Llama 4. A Dutch company may not run Llama 4 itself to build that service.

Meta's new open model, Muse Spark Open from 15 October 2026, has no EU exclusion. Its licence turns on two other conditions: a separate agreement as soon as you offer the model to others as a service and your group has more than $20 million in revenue, and the mandatory notice "Built with Muse Spark Open" in the user interface of every product, according to the licence. Internal use falls outside the first condition.

Which revenue thresholds sit in open-weight licences?

Revenue thresholds in open-weight licences almost always concern the revenue of the whole group, not of the entity using the model. That matters for European subsidiaries of international groups.

The strictest for our readers is Mistral's Modified MIT licence, which covers Mistral Medium 3.5 and Devstral 2. The licence text says: "You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million". That's about $240 million a year. A standalone company of 300 people is usually below it. A 300-person subsidiary of a billion-euro group isn't, and may not use the model at all, internally included.

The thresholds in Kimi K3, Muse Spark Open and the large Qwen models work differently: they apply only when you offer the model to third parties as a service, and internal use is exempt. If you build your own product in which customers talk to the model directly, take care. On top of that, Kimi K3 and Qwen require the model name to be visible in your product above 100 million users or $20 million in monthly revenue.

Work this out before you invest. Building on a licence you lose at the next acquisition or growth step is a risk that also belongs in your contract terms for AI software.

Is an open-weight model the same as open source?

An open-weight model isn't the same as open source, and the difference is more than semantics. The Open Source Initiative published the Open Source AI Definition 1.0 on 28 October 2024, which requires, besides the weights, "the complete source code used to train and run the system" and enough information about the training data. Hardly any major model provides that. They're open-weight models.

On Llama, the OSI is blunt: the licence is "still not Open Source by any stretch of the imagination", the OSI wrote in February 2025, partly because it discriminates against users and excludes "any persons in the European Union".

For a business, the practical question is: may I use, modify and distribute the model without asking permission? Under Apache 2.0 and MIT the answer is yes. Under custom licences it depends on who you are and what you do with it.

What does the licence mean under the AI Act?

Under the AI Act, the licence decides whether the model's maker is exempt from part of its obligations, and that has a consequence for you. Article 53(2) exempts providers of models under "a free and open-source licence that allows for the access, usage, modification, and distribution of the model" from keeping technical documentation and supplying it to companies that build the model into their systems. The exemption doesn't apply to models with systemic risk.

Which licences count as free and open is set out in the European Commission's guidelines of 18 July 2025. Not free, according to those guidelines, are among other things "usage restrictions triggered by user scale thresholds" and "requirements to obtain separate commercial licences for specific use cases". A licence that requires payment above a certain scale counts as monetisation. Attribution requirements are allowed.

The last column is our reading of the guidelines, not a Commission ruling on any specific licence. The practical point: under a free licence the maker has no legal duty to help you with documentation, so you'll need to get that information from the model card yourself.

You only become a model provider yourself if you modify it substantially, according to the guidelines at more than a third of the original training compute. Ordinary fine-tuning by a company stays far below that, as we explain in our piece on AI model distillation. That piece also covers the question left out here: whether you may use a model's output to train another model.

How do you choose a model on licence?

You choose a model on licence with four questions, in this order. A no on either of the first two rules the model out, however well it scores.

1. Does your company have its principal place of business in the EU? Then Llama 4 is out.
2. Is your group's consolidated revenue above $20 million a month? Then Mistral Medium 3.5 and Devstral 2 are out.
3. Do you offer the model as a service to customers or third parties? Then you may need a separate agreement for Muse Spark Open, Kimi K3 and the large Qwen models. Not for internal use.
4. May the model's name be visible in your product? Muse Spark Open always requires it, Kimi and Qwen above a threshold, MiniMax M3 for any commercial use.

To avoid all of these questions, choose a model under Apache 2.0 or MIT. That's no longer a constraint: Gemma 4, gpt-oss, Mistral Small 4, DeepSeek V4.1 and the 27-billion-parameter Qwen models together cover almost every business task you'd want to run locally. How far they trail the most expensive closed models is in open-weight models versus the frontier. What else is at stake with the Chinese models in this table, beyond the licence, is covered in Chinese AI models at work, and we'll update the licence for Mistral Large 4 as soon as Mistral publishes it.

The full route from model choice to operations is in our guide to local AI for business. Want an open model under your own control, on a licence that still holds after your next growth step? We build that as part of local AI.