Mistral Large 4 explained: what Mistral's largest model is actually for
Mistral Large 4 went into preview on 6 October 2026. Our view on what it is good for, where it trails, and why its languages matter most.
More than 160 languages. That figure appears once in Mistral's launch post for Large 4, at the end of a paragraph about where the model will run, and it is introduced with two words: "Fun fact". The full sentence reads: "Fun fact: a significant share of ML4’s training data was multilingual, spanning more than 160 languages, including every official language of the European Union."
We think that sentence deserves better billing. Much of the post, published on 6 October 2026, is about cybersecurity, software, finance and law. That tells you who Mistral hopes to sell to first. For someone in Lisbon, Ghent or Tallinn who wants help with ordinary life in their own language, the line about languages is the headline. We are a small Belgian company, and Belgium works in Dutch, French and German, so the question of which languages a model handles well comes up at our own desks.
What Mistral has released so far
As of 19 October 2026, Large 4 is a public preview. Mistral calls it "our largest and most capable model to date", and the launch post gives its official nickname, with a straight face, as "le Chonk". It is built as a mixture of experts: about a trillion parameters in total, with roughly 50 billion of them doing the work for each piece of text it writes. It reads images as well as text, and Mistral's own examples lean on documents, charts and photographs.
Two promises remain open. The first is the weights. Mistral says it will release them "by the end of the month", and the model's page on Hugging Face, which still read "Upcoming release" when we checked on 19 October, gives an expected date of 31 October 2026. Mistral's own models page already labels Large 4 "Open-weight". We will wait for the files to be downloadable before using that word, because that is what open weights means. Mistral has not yet said which licence they will come under.
The second promise concerns location. Mistral says Large 4 will be offered in several regions, "including a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law." The company also says the model was trained in its own data centres in Europe and that the preview runs on that same infrastructure.
We added Large 4 to MultiChats on 12 October 2026, on paid plans. A message to it goes straight to Mistral's own service, with no router in between, and our privacy policy lists Mistral as an EU-based provider. This piece stops there on data. What "European" means for your conversations more broadly has its own article.
The case against it, made properly
The strongest objection comes with numbers. Artificial Analysis, an independent group that runs the same battery of tests across models, scores the Large 4 preview at 38 on its Intelligence Index (checked on 19 October 2026). On the same index, Claude Opus 5.5 scores 58, and Google's Gemini 4 Argon and OpenAI's GPT-6.1 Sol both sit in the low fifties. The same page describes Large 4 as "slower than average and very verbose", and records it producing more than twice the median amount of text over the test run.
If you have one hard question, a tangled decision or a piece of reasoning where being wrong costs you something, the model twenty points higher on an independent index is the better bet. A slower, wordier reply makes that gap more noticeable. Pride in a European model does not change a test score.
That argument is right as far as it goes. We accept its conclusion: today, we would send our hardest reasoning questions elsewhere. Two points qualify that judgement. The scores belong to a preview, and Mistral says the training run behind it "is still in flight", with "large and rapid improvements in the weeks and months to come" expected. That is a promise, and we will believe it when the numbers move. Wordiness is common near the top too: the same index calls Claude Opus 5.5 "very verbose" as well.
Where we land
Much of what people bring to a chat assistant is ordinary and heavy on language. A reply to a landlord in Dutch. A cover letter in Portuguese that should sound as if a native wrote it. A photo of a Greek form that needs explaining before Friday. The index that ranks Large 4 at 38 is built from tests of office work, coding, science and general knowledge, and it says very little about whether a model gets the tone right in Estonian.
For European readers, the case for Large 4 rests on its languages and on where it runs, much more than on where it sits in a ranking. It is worth trying as the model you write and read with in your own language, especially if that language is one of the smaller ones in the Union, and especially if it matters to you that the company behind it trains and serves it in Europe. Keep your hardest questions for the models at the top of the index.
One limit matters. We have not seen a per-language comparison we would stake a recommendation on, so use Mistral's language claim as a reason to test it on your own writing. Give it a paragraph you wrote in your language and ask what a native speaker would change and why. Our guide for people whose first language is not English turns that into a two-minute check you can run on any model.
In MultiChats, Large 4 accepts images, so a photograph of a letter or a screenshot of a chart works. It does not take PDFs directly: with Large 4 selected, the app will tell you to switch to a model that accepts them. A Word document does go through, because we pull the text out before it reaches the model, and our piece on which AI handles the longest documents covers the longer cases. It cannot search the web, so questions about this week's news belong on a model that can. It thinks before answering by default; you can switch this on or off, with nothing in between. By default, the website decides for you. Turn off Intent Detection in Settings, under Tool Preferences, to get the control back; the phone apps always show it.
Mistral Large 3 is still in the picker too. The practical difference is that Large 4 can think before it answers and Large 3 replies straight away; both read images.
We will come back to this piece when the weights arrive, and again when the preview label comes off and the scores are run on the finished model.