Mistral launches Large 4, a 1-trillion-parameter model it will release with open weights

On Tuesday, Oct. 6, 2026, Mistral opened a public preview of Mistral Large 4, a 1-trillion-parameter model that reads text and images. Mistral says it is the strongest open-weight model built in the US or Europe and leads on cybersecurity work, and it plans to release the weights by the end of the month.

Most of the best AI models are closed, and most of the best open ones come from China. Mistral is pitching Large 4 as the open option from the US-and-Europe side that companies and governments can run on their own machines. The cyber angle is the sharp part: Mistral says its model will do security work that closed models refuse, which is useful for defenders but is also exactly why it's holding the weights back for a few weeks of testing. The benchmark numbers are Mistral's own, so the real test comes when the weights are out and outside researchers can check them.

On Tuesday, 6 October 2026, Mistral published “Introducing Mistral Large 4.” The page prints October 6, 2026. It does not print an hour. Mistral says it is opening a public preview of Mistral Large 4. The nickname on the page is “le Chonk.” Mistral also shortens the name to ML4. The preview is available now, through an API on Mistral Studio. An API is the door a program uses to send the model a request. Mistral says the weights will come by the end of the month. Weights are the learned numbers that make the model what it is. Releasing them means someone else can run the model, not only call it through Mistral. Until then, Mistral says it is red-teaming the model with cybersecurity leaders, vetted partners, and state authorities. Red-teaming means testers try to make the model fail, or do harm, before a wider release. Those groups get a version with reduced moderation and expanded cyber capabilities. Those lines are Mistral’s. The page does not name a price, a license for the weights, or a calendar day inside the month.

The size, as Mistral states it. Large 4 is a 1-trillion-parameter model, and Mistral’s largest so far. A parameter is one learned number inside the model. Mistral says 49 billion of those parameters are active. Active means the slice that runs for a given request. The model is natively multimodal. That means it was built to read text and images together, not as a text model with pictures added later. Mistral says a significant share of the training data covered more than 160 languages, including every official language of the European Union. Those lines are Mistral’s.

Where Mistral says it built and serves the model. It trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs, in data centers Mistral runs in Europe. A GPU is a chip built for the heavy math these models need. From scratch means the lab did not start from another company’s finished model. The public preview runs on that same hardware. Mistral says the model will be available in several regions, including a European deployment it operates end to end, on its own, under European law. Those lines are Mistral’s.

What Mistral claims the model can do, relative to other models. It says Large 4 is already competitive with the strongest open models globally, and that it significantly outperforms any open-weight model developed in the United States or Europe. Open-weight means the learned numbers are meant to be released so someone else can run them. On cybersecurity, finance, and law, Mistral says Large 4 is state of the art among open models. State of the art, in that sentence, is Mistral’s claim that it leads that group. Those lines are Mistral’s. They are not a score an outside lab has published as its own.

The security scores, as Mistral reports them. Mistral cites the Artificial Analysis Cyber Index, which it describes as an independent test of how well models find and fix security flaws in real software. Mistral says Large 4 ranks in the top five on that index. One test in the index asks a model to reproduce a real vulnerability in open-source software and then patch it. A vulnerability is a flaw someone could use to break in. A patch is the fix. Mistral says Large 4 scores 82 percent on that test, and that this is the highest score of any model. Mistral also says the model solves 93 percent of the 40 challenges in Cybench, exercises drawn from security competitions, and calls that one of the highest scores reported for an open-weight model. Those figures are Mistral’s.

Mistral’s account of models that refuse the same work. It says several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on that reproduce-and-patch test because they refuse the task. Closed means the lab keeps the weights, and a customer uses the model through the company’s own service. Mistral’s point is that defending software often starts by proving a flaw is real, and that the safety filters in closed models can block that step. Those lines are Mistral’s.

The coding scores, as Mistral reports them, using numbers it says come from Artificial Analysis. Large 4 scores 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA, and 28.3 percent on Terminal-Bench 4. Those are scored tests of software work. A terminal is the text window where a person types commands. Mistral also says it ran a blind coding evaluation with Surge AI. Professional raters scored the answers from 1 to 5, and the model names were hidden. Mistral says Large 4 ranked second of five models, at 3.74, behind only Claude Opus 5 at 4.22. On that same list, Mistral prints Kimi K3 at 3.59, GLM-5.3 at 3.60, and GLM-5.2 at 3.40. Those figures are Mistral’s account of the evaluation.

The agent score, as Mistral reports it. An agent, here, is software that can take steps in other apps, not only answer in a chat. Mistral says Large 4 scores 59.9 percent on AutomationBench, a set of 657 business workflows across apps such as Gmail, Google Sheets, Slack, and Salesforce. That score is Mistral’s.

Reuters reported the same Tuesday from the Ai Everything conference in Abu Dhabi. Chief executive Arthur Mensch told the room: “The model we're actually announcing today is actually above the Chinese models on certain aspects, including cyber.” He did not name the Chinese models, and he did not name the measure. Reuters says Mistral raised €3 billion at a €21 billion valuation in a funding round in September, and that the company last launched a model in April. Those lines are Reuters’. Mistral’s own post calls that September round a €3 billion Series D, and calls it the largest equity round ever raised by a European technology company. That sentence is Mistral’s.

CNBC, by Kai Nicol-Schwarz, reported the unveiling the same day. It says Mistral called Large 4 the strongest open-weight model developed outside China by a “substantial margin,” and that the preview goes to developers and cybersecurity leaders, alongside state authorities, before a wider release later this month. CNBC also writes that the model still lags the frontier on coding. That lag line is CNBC’s. Guillaume Lample, Mistral’s co-founder and chief scientist, told CNBC the model will keep improving as the company adds training capacity after the Series D, and that the cyber tools are meant to help companies and governments defend against people who jailbreak closed models in order to attack. A jailbreak, here, means getting around a model’s safety limits. Those sentences are CNBC’s account.

The picture is a cream card. The title reads Mistral Large 4, with the nickname le Chonk under it. Four lines list the size, 1 trillion parameters with 49 billion active, a preview API today, open weights by the end of October, and training on 3,800 Nvidia Grace Blackwell GPUs. A panel beside that list is labeled Cyber. It shows 93 percent on Cybench and 82 percent on a reproduce-and-patch test. The bars run from orange to red. The frame does not print a calendar date. It is a summary of the announcement. It is not a photograph of a lab, a chip, or a person.

In plain terms, Mistral on Tuesday opened a public preview of its biggest model, a system of 1 trillion parameters that it nicknames le Chonk, and said the weights will come by the end of the month. The company says it leads the open models built in the United States and Europe, and that on a security test closed models refuse, this one scores 82 percent. Those benchmark numbers are Mistral’s. The weights are not public yet.