What does it mean that Mistral Large 4 is open weight?
Picture a clinic's records office, or a town hall's registry desk. It wants an AI assistant, but it cannot send a single file to an outside server. On 29 October that office got a new option. Mistral, the French AI company, made the weights of Mistral Large 4 public. Open weight means you can download the model and install it on your own servers, instead of using it only through someone else's service.
The model had already been on show in a public preview earlier in the month, OrcaRouter notes. What changes now is who runs it. TechCrunch reports that businesses and governments can deploy the system themselves. Mistral says it trained the model from scratch on European infrastructure. Its page on Hugging Face lists more than 160 languages, including every official EU language.
What can you do with it, and under which licence?
Mistral's licence is generous. You can use, modify and adapt the model for free, and build commercial products on it. There is one condition. A business that sells direct access to the model as a service needs a separate commercial licence. According to OrcaRouter, which analysed the text, internal use and products that use the model inside a feature fall outside that condition.
Once the licence is clear, the next question is where the model runs. Today it can reach you in three ways.
- Pay per use through Mistral's own service, the API. There is nothing to install, but your requests travel to Mistral.
- A private deployment in a cloud you already use. Mistral says the model is available through Microsoft Foundry, with private deployment options.
- In-house. You download the model and run it on your own servers. Your documents stay inside your perimeter, and keeping it running becomes your job.
Does running your own AI model save money?
Usually not, and it is worth saying plainly. IBM's chief executive, Arvind Krishna, told investors that open weight models now deliver about 90% of the effectiveness of the best closed models. According to The Register, they do it at 10% to 25% of the cost per use. The hard part, he added, is finding people who can run them.
That is where the bill starts. OrcaRouter counts a download of about 1 TB, and The Next Web reports that even the compressed version needs four top-end Nvidia graphics processors. For almost every team, OrcaRouter concludes, running it in-house costs more than the discounted API. What you buy is control over where the data sits and how the model is adapted. Today's API price is also a launch promotion with no published end date. This may change; we will update this page.
Why does it matter for your business or public office?
It still matters, because in many offices the obstacle is the data, not the AI. Think of patient files in a clinic, applications at a municipal desk, contracts in a law firm, payslips in HR. Until now, a top model meant sending those documents to a provider. With an open weight model, the summary of a medical report or the draft of a registry entry can be produced on a machine you control.
That platform handles healthcare laboratory data, yet it runs on Google Cloud, not on in-house servers. Our read: control does not come only from where the server sits. It comes from who can see what, what gets logged and who approves the sensitive steps. In-house is one way to get there, and a well-governed cloud is another.
What about us?
That is also how we build. Our systems are model-agnostic. Claude, ChatGPT, local models or our own agents sit behind the same gate, and you can switch models without touching the system. They run in your perimeter, on your cloud or on your servers. So a model like Mistral Large 4 is one more option behind that gate, not a new project.
How we would use it: take a public office that registers incoming documents every day, the job our document protocol agent does. The model would run on the office's servers. It would read each document, propose the classification and the entry, and log every step. The signature stays with the official. If a better model arrives next month, you swap it without rebuilding the workflow. This is the work of our AI agents and automation service.
So should you run your own model?
- If your data cannot leave your perimeter (health records, public files, legal documents) and you handle it every day, run the model in-house or in a private deployment.
- If your data is ordinary and you want results this month, use the API. It is the cheaper way to start.
- If you are not sure, test the model through the API on a real, non-sensitive example. Build the system so you can move it in-house later without rewriting it.
“Where the model runs should be a decision you can change. Build the workflow first, so the model can move from the API to your servers, or from one provider to the next.”
DBrothers — how we scope systems
What else should you know before you decide?
Is Mistral Large 4 free?
The licence is free for internal use and for products built on it. The hardware, the energy and the people to run it are not.
Can a small company run it in-house?
Not on an office computer. Even the compressed version needs server-class hardware. The Next Web adds that Mistral plans smaller versions, including one for laptops, with no dates yet.
Is it the best model for my work?
That depends on your documents, not on a leaderboard. The published scores are Mistral's own, and OrcaRouter notes that no independent check had come out yet. Test it on your own examples before you choose.
Leaderboards change every month; your team's work doesn't. If you want to know whether Mistral Large 4 handles your documents, and whether it belongs in the cloud or on your servers, start by trying it on one real example with us.
