French AI company Mistral released downloadable weights for an updated version of its Mistral Large 4 model on October 29, Reuters reported. The release lets organizations run and adapt the model on their own infrastructure, but businesses offering it as a hosted service need a separate commercial license.

“Downloadable” does not mean unrestricted, or easy to run. The license draws a line between using the model in your own products and giving customers direct access to it as a service.

A large model with substantial hardware needs

Mistral Large 4 has about 1 trillion parameters, with 49 billion active for each token, according to Hugging Face’s model overview. Hugging Face’s deployment guide describes a setup using eight Nvidia H200 graphics processors. Downloading the files, then, is not the same as running the model on an ordinary laptop.

Self-hosting can give organizations more control over where their data goes and how they adapt the model. Mistral says additional training and testing with outside users improved coding and defensive cybersecurity.

Where the license draws the line

Mistral’s modified-MIT-style license allows users to use, modify, fine-tune and commercialize the model. But businesses that offer direct access to it as a “Model-as-a-Service” must obtain a separate commercial license from Mistral.

The license excludes internal tools and products that use the model within specific features from that service requirement. Businesses planning to give customers direct access through a hosted endpoint or application programming interface should check the terms carefully.

That distinction matters in practice: a company can build a product with the model inside it, but a service that lets customers query the model directly falls under the separate licensing requirement.

What the performance figures show

Mistral says additional training improved coding and defensive cybersecurity. Its release materials report these benchmark results:

  • A 13-percentage-point gain over the preview version on Terminal-Bench 4.0, a test of terminal-based coding tasks.
  • A 51% score on the Artificial Analysis Cyber Index, which tests whether a model can find, reproduce and validate software vulnerabilities, then patch them without breaking existing functionality.

These are results on defined tests, not guarantees of performance in every company’s coding or security environment.

Mistral Large 4’s weights are available to download and adapt. But businesses offering customers direct hosted access need separate commercial permission from Mistral.