Mistral published the weights. Read the model-as-a-service line.

On October 29, Mistral released the weights of an improved Mistral Large 4, its 1-trillion-parameter, natively multimodal model. The company says it had offered that model in a public preview earlier this month. Mistral says an organization can use and improve the model without depending on Mistral to serve it through an API. If you are deciding whether to call a vendor, run a model yourself, or hire someone to run it, see Consultant, agency, build team or hire?.

What shipped

The announcement sets the API price, starting that day, at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. Those rates are company-reported. The same post says the weights fall under the Mistral Ecosystem Open-Weight License, which Mistral calls a modified MIT license, and that a business may use, modify, fine-tune, and commercialize the model, with a condition when the business offers the model as a service.

Mistral’s company-reported figures on the Artificial Analysis Coding Agent Index are a score of 48 percent, described as effectively tied with GLM-5.3, and 47 percent fewer tokens per task on average. A lower token count changes the API bill. The owner still picks the process.

The model has three reasoning modes. Low is for simple tasks, and it replaces a mode Mistral called None. Medium is for everyday tasks. High is the default, for the hardest work. The announcement publishes one set of API rates and names High as the default.

The license line

The announcement’s condition is one sentence. The license it links to is the text to mark a product against. What follows describes that file.

The license grants rights to use, copy, modify, distribute, and run the model, including further training and derivative works, and then states conditions. Section 2 says that if the licensee, or any affiliate, operates a model-as-a-service business, the licensee must enter a separate commercial license with Mistral before using the model or a derivative for any commercial purpose.

The license defines that business as giving a third party access to the model, or a derivative, for inference or post-training by any means. An API or a hosted endpoint is the example it gives. That access has to be in a manner that lets the third party exercise meaningful control over the inputs, the parameters, or the training data. The definition sets two cases aside: an end-user product whose model capabilities sit only inside specific features or harnesses, and mere relaying of requests to models hosted by others.

Section 3 places two uses outside the section 2 requirement. One is internal use, defined as any use that does not make the model or a derivative available to third parties. The other is any use through Mistral’s official products.

Hold the two sections together. Section 2 names the commercial-license trigger, including when an affiliate is the one operating that kind of service. Section 3 names the uses placed outside that trigger. Mark your product in those words. The shorter sentence on the blog is a pointer to the file.

Price a week on paper

These counts are round placeholders so the rates are easy to check. They are not a customer’s usage. Count each token once.

• 5 million uncached input tokens at the company-reported $0.70 per million come to $3.50.
• 15 million cached input tokens at $0.07 per million come to $1.05.
• 2 million output tokens at $2.10 per million come to $4.20.

The placeholder week comes to $8.75 on the API. Swap in a real week from your own logs before you compare paths.

That figure is the API path, at the rates Mistral posted. A path on your own machines still needs a number for machines, power, and the person who keeps the process up. Collect that number from your own quotes. The October 29 post supplies the API side.

What stays with the owner

The weights and the rate leave three lines for the owner to write.

1. The process allowed to depend on this model.
2. Whether that process calls Mistral’s API or runs on machines the owner controls.
3. The person who checks the output before it goes out. The company-reported coding score is a benchmark line. The check is a name.

What to do this week

• Name one process you would move onto this model. Write what happens to the business on a day the API does not answer.
• Run the three rates against a real week of that process. Keep the placeholder total, $8.75, in the margin, and write your own total beside it.
• Open the Mistral Ecosystem Open-Weight License and mark one category in its words: internal use, embedded in specific features or harnesses, relaying to a host you do not run, or third-party access with meaningful control over inputs, parameters, or training data. If an affiliate offers that access, add the affiliate on the same line.
• When the mark is third-party access with meaningful control, copy section 2 onto that page. Section 2 requires the licensee to enter a separate commercial license before commercial use when the licensee or an affiliate operates that kind of business.
• When the plan is to run the weights, write the operator’s name on the page. The posted API rate is the plan you can cost from the announcement alone.

Which process stays on the posted rate, which one moves onto machines you control, and who operates it is the kind of decision we make with owners in a consult block. The price card will move. The question of who can keep the work running stays.