How do you try Mistral Large 4?

Mistral Large 4 is the new flagship model from Mistral AI, released as a public preview on October 6. Mistral describes it as a model of about one trillion parameters, the internal numbers a model learns during training, of which 52 billion are active for any given piece of text. That design, called a mixture of experts, lets a very large model run at the cost of a much smaller one. It reads text and images and writes text.

Two words in the announcement matter for anyone trying it now. "Preview" means the model is still being trained: Mistral says its reinforcement learning run is ongoing and expects quick improvements over the coming weeks. "Open weights" means the trained model files will be downloadable so organizations can run it on their own machines; Mistral says that release is planned for the end of this month. Until then, you try it on Mistral's own service.

Where to try it

Mistral's model page lists a context window of one million tokens. A token is a small piece of text, roughly part of a word, and the context window is how much text the model can consider at once. The listed preview price is $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14. Mistral's model page also lists support for function calling, structured outputs and questions about documents.

Mistral serves the preview from its own data centers in Europe, including a deployment it runs under European law. For some organizations, where the data is processed matters as much as the score.

A worked bilingual test

Launch charts tell you how a model does on someone else's tasks. A personal test tells you how it does on yours. This one takes about twenty minutes and needs nothing more than a chat window.

Step 1. Pick five tasks you really do. Write each once in English and once in French. Keep the content identical.

Step 2. Run all ten prompts in fresh conversations, so earlier answers don't influence later ones.

Step 3. Score each answer from 1 to 5 on accuracy, tone and correct language, and note anything you would have to fix.

Here is one pair as an example.

Write a short, polite email to a client explaining that their website
update will be delivered next Tuesday instead of Friday, because we are
waiting for their photos. Keep it under 120 words. Use a warm,
professional tone.

Rédige un court courriel poli à un client pour lui expliquer que la mise
à jour de son site Web sera livrée mardi prochain plutôt que vendredi,
parce que nous attendons ses photos. Moins de 120 mots. Ton chaleureux
et professionnel, en français du Québec.

A score sheet could look like this. The numbers are illustrative, to show the format.

The pattern is what you are looking for. If French scores trail English on the same tasks, you know where to check its work closely.

What to keep in mind with a preview

Because training is still running, results can shift from week to week. Save your prompts and your scores, and rerun them after Mistral announces an update. Mistral's headline benchmark claims are its own: it reports 61.7% on DeepSWE v1.1, a test of long coding tasks, and calls the model the strongest open model in cybersecurity, finance and law. Those claims will mean more once outside testers have the open weights.

How does it handle French?

Mistral says a large part of Mistral Large 4's training data was multilingual, spread across more than 160 languages, among them all of the EU's official ones. French is one of those, so on paper it should be well covered.

Mistral has not published a separate French benchmark score in its launch post, so your own test matters. Three things are worth checking. First, regional usage: European and Canadian French differ in vocabulary ("courriel" or "e-mail," "fin de semaine" or "week-end"), so say which one you want, as in the prompt above, and see whether it follows. Second, register: ask for both formal "vous" and friendly "tu" versions of the same message and check that the model keeps the choice consistent. Third, accuracy under pressure: give it a longer French document to summarize, then check names, numbers and dates against the original.

If the French answers hold up on your own tasks, that is a stronger signal than any chart. If they slip, give the model an example of the style you want at the top of your prompt; a sample paragraph often fixes tone faster than a list of rules.