Mistral Large 4’s licence puts the business model in the small print

Mistral Large 4’s licence reportedly allows commercial use without a revenue ceiling, but puts a separate licensing condition on some model-as-a-service businesses. That distinction could matter more to developers choosing how to deploy the model than the word “open” on its own.

Watch Desk analysis

What happened

Forkast says Mistral has replaced a revenue-based threshold with a condition tied to how a business offers access to the model. Under the account, a company operating a model-as-a-service business must sign a separate commercial licence before using the model or its derivatives commercially. The article describes that business as giving third parties access to inference or post-training while allowing them meaningful control over inputs, parameters or training data.

The report also describes exemptions: products with model capabilities embedded in specific features, internal use, use through Mistral’s own products, and simply relaying requests to models hosted elsewhere. Forkast says other labs in the comparison, including StepFun, Meta and Moonshot AI, retain revenue-based thresholds. Those are the article’s descriptions of the terms, so developers should check the actual licence against their own service before relying on the distinction.

Why it matters

A revenue ceiling is relatively easy to spot. A condition based on the service a company provides asks a different question: does the customer get meaningful control over the model’s inputs, parameters or training data? That could leave an embedded feature on one side of the line and a hosted model service on the other, even if both are commercial products.

For teams weighing open-weight models, the practical lesson is to examine the delivery model as well as the download. The licence could affect whether a business can offer direct model access, while leaving other commercial uses available under its stated carve-outs.

Our read

This is a more useful licensing debate than a tidy “open versus closed” label. A model can be commercially usable and still place a condition on a particular kind of service. Forkast’s account gives developers a map of the reported distinction, not a substitute for reading the licence or getting advice for a specific deployment. The small print, as ever, has declined to stay small.

What to watch

• Whether Mistral publishes clearer examples of what counts as meaningful customer control.
• How developers and hosting providers interpret the model-as-a-service condition in practice.
• Whether other labs change their revenue thresholds or retain them.

Discussion spark: Should open-weight licences be allowed to distinguish between embedded product features and hosted model services, or does that make commercial reuse too hard to judge?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.