Mistral Large 4 shows up a summer late

In July, Mistral AI's CEO Arthur Mensch promised "a very exciting model for this summer."¹ Summer ended on September 22. The model arrived on October 6, on the stage of the AI Everything conference in Abu Dhabi.² It's called Mistral Large 4, and the company's unofficial nickname for it is "Le Chonk," the internet's word for a cat that has been eating well.³ For a model with a trillion parameters, that's just accurate labeling.

Mistral Large 4 is the most capable open model built outside China, according to the independent rankings of Artificial Analysis, which score it at 38 points.⁴ It still trails seven Chinese open models and sits 20 points below Claude Opus 5.5.⁵ Developers have been able to try it through the API, the access point meant for software, since October 6, as a preview. Its weights, the files that let anyone run it on their own hardware, come out on October 27.²

The summer model

In September, when Mistral raised €3 billion, I noted that the money would go into concrete first, meaning data centers, and that the frontier model, the kind that competes with the best of the moment, would have to wait. It waited a month. Mistral calls Large 4 "the first milestone" on the roadmap that round paid for.³ It also came out of that concrete, since it was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in the company's own European data centers.³

Parameters are the internal settings a model tunes during training, and their count gives a rough sense of its size. Large 4 has a trillion of them, but it only uses 49 billion for each word it produces.⁶ That's how a mixture of experts works. The model hands each piece of text to a few specialized subnetworks instead of the whole thing, which makes it far cheaper to run than to store. It reads text and images, and it learned more than 160 languages, including every official language of the European Union.³

The venue tells you the rest. In Abu Dhabi, Mensch pitched Mistral's independence from both the United States and China, a selling point he says lands well in the Gulf and across Asia-Pacific.² Europe's sovereignty model, in other words, made its debut in the Emirates, in front of customers who are mostly shopping for neutrality.

First in the West, eighth on the chart

Artificial Analysis is an independent firm that puts every model through the same battery of tests and boils the results down to a single score. Large 4 gets 38 points, up from 14 for Mistral Medium 3.5 and 9 for Large 3, the December 2025 model.⁷ That's a huge jump, and the firm concludes that France once again has "the most intelligent model from outside the US and China."⁴

The rest of the chart is less festive.

Seven Chinese open models rank higher, with Xiaomi on top.⁷ Among closed models, Large 4 ties GPT-6 Luna, the small model OpenAI shipped alongside Sol in September.⁴ In Abu Dhabi, Mensch said Large 4 was "above the Chinese models on certain aspects, including cyber," without saying which models or how he measured it.² Mistral's own charts are more modest than its CEO. In a blind evaluation where human testers graded code, Large 4 scores 3.74 out of 5, behind Claude Opus 5 at 4.22,³ and that isn't even the current Opus.

The list price is $1.36 per million tokens⁸ in and $4.18 out, cut in half for the first two weeks.⁴ That's within a few cents of GLM-5.2, the Chinese model Mistral has been serving on its own platform since August, at $1.40 and $4.40.⁹ The French model and the Chinese one now carry the same price tag on the same shelf. The price per token still doesn't tell you the bill, and Artificial Analysis did the math. Per task in its test suite, Large 4 costs $1.13 at full price, against $0.25 for GLM-5.3 Flash, which also scores higher.⁴

The model that will look at your bugs

Mistral leads with cybersecurity. On a test that asks a model to reproduce a software vulnerability and then patch it, Large 4 succeeds 82% of the time. According to Mistral, Claude Opus 5.5 and GPT-6 Astra score close to zero because they refuse to do the exercise.³ The big American labs keep their most offensive models under lock and key, reserved for a vetted circle, and their public versions turn down anything that looks like an attack. A defender trying to test their own flaw hits the same wall as an attacker.

Mistral isn't throwing the doors open either. The company says its model refuses malicious requests more often than any other open model.⁵ During the preview, a build with lighter moderation goes to security firms, vetted partners, and government agencies, whose job is to attack it before the weights are published.⁵ This is called red teaming, which means inviting people to break your model before others do it uninvited.

So Mistral holds a key of its own, for three weeks. On October 27, the weights go public, and the key goes with them. Last month I wrote about Heretic, a free tool that strips the censorship from an open model, and every refusal along with it. I doubt Large 4's carefully calibrated refusals will survive the download for long. They will survive in Mistral's API, though, and I find that bet more coherent than the American one, which keeps the key from everyone. The compressed versions of the model will fit on four to eight GPUs,¹⁰ which is slightly more than the one in my gaming laptop.

The gap changes units

In September, I wrote that I would switch to Mistral the day the gap with Claude was measured in months, and that it was measured in generations. Large 3 had 9 points while Opus 5.5 has 58. Large 4 has 38, so the gap now comes down to one step, the one that separates OpenAI's entry-level model from the ones above it.

For my prototypes, the mockups that get business teams reacting before anyone writes a spec, I'm staying on Claude, because at that level of demand, a step shows right away. For a well-bounded task, like sorting requests or pulling fields out of a document, Large 4 becomes a serious candidate. That holds especially where data isn't allowed to leave the building, since open weights mean you can host it yourself.¹⁰ Mistral also says it will become the default model in Vibe, its consumer assistant, formerly Le Chat.¹⁰ That's the door I walked out of.

The next date is October 27, with the weights, a license Mistral hasn't officially said anything about yet,⁷ and the first tests run by people who don't work for Mistral. Until then, I'm holding on to one number, and for once it flatters Mistral. In a single model, the company cut its deficit to Claude from 49 points to 20.

Notes

1. Tech Times, "Mistral AI Targets Frontier Gap With Open-Weight Model Entering July Early Access", July 6, 2026. ↩

2. Reuters, via Yahoo Finance, "Mistral CEO says new AI model beats Chinese ones in some areas", October 6, 2026. ↩ ↩² ↩³ ↩⁴

3. Mistral AI, "Introducing Mistral Large 4", October 6, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

4. Artificial Analysis, "Mistral has released Mistral Large 4, making France home to the most intelligent model outside the US and China", October 6, 2026. The Opus 5.5 and GPT-6 Luna scores are for their maximum reasoning setting. ↩ ↩² ↩³ ↩⁴ ↩⁵

5. The Decoder, "Mistral Large 4 is Europe's trillion-parameter answer to US models that refuse security work", October 6, 2026. ↩ ↩² ↩³

6. Those are the announcement's numbers. The documentation published the same day says 1.05 trillion parameters with 52 billion active and a working memory of a million tokens, while Artificial Analysis counts half that (MIXED). Three billion parameters is roughly the size of a model that runs on a phone, misplaced somewhere between two pages of the same website. ↩

7. Trending Topics, "Mistral Large 4 on Artificial Analysis: Eighth Among Open Models", October 6, 2026. ↩ ↩² ↩³

8. A token is the unit of text these services bill by, roughly a short word or a piece of a longer one. ↩

9. Mistral AI, Z.ai GLM 5.2 model documentation. ↩

10. Blog du Modérateur, "Mistral Large 4 : Mistral AI veut réduire l'écart avec les modèles open weight chinois" (in French), October 6, 2026. ↩ ↩² ↩³