Mistral launches a 1 trillion parameter model called Le Chonk and says the open weights land at the end of October: Reveneau

Mistral launches a 1 trillion parameter model called Le Chonk and says the open weights land at the end of October

Mistral released a public preview of Mistral Large 4 on 6 October. The 1 trillion parameter mixture-of-experts model has 49 billion active parameters, was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European datacenters, and Mistral says it will publish the weights by the end of October after red-teaming is finished.

Why it mattersA team that cannot or will not send code to a closed model has a new open-weight option in three weeks, and Mistral says it leads outside China on cyber work that American and Chinese closed models often refuse to perform.

A team that cannot or will not send its code to a closed model has a 1 trillion parameter open-weight option in three weeks. Mistral AI released a public preview of Mistral Large 4 on 6 October, nicknamed Le Chonk, and said the weights will be published by the end of the month after red-teaming is finished.

The model is a mixture-of-experts design with 1 trillion total parameters and 49 billion active per token, and is natively multimodal. Mistral says Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own datacenters in Europe, and that the public preview is served from the same hardware. Training data covered more than 160 languages, including every official language of the European Union, Mistral says.

Where Mistral says it leads

Mistral's own numbers for Large 4 put it among open-weight models at the top end. The company reports a 61.7 percent score on DeepSWE v1.1 for software engineering, 59.4 percent on SWE-Atlas-QnA, and 28.3 percent on Terminal-Bench 4. On AutomationBench, a set of 657 business-tool workflows across Gmail, Google Sheets, Slack and Salesforce, Mistral reports Large 4 at 59.9 percent, ahead of Kimi K3 and DeepSeek V4 Pro by its own measurement.

On cybersecurity, Mistral says Large 4 scores 82 percent on CyberGym-E2E, a test on the Artificial Analysis Cyber Index that asks a model to reproduce a vulnerability in open source software and then patch it, and 93 percent on Cybench, a set of 40 security-competition tasks. Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero on the first of those tests because they refuse to perform the task. These are Mistral's own numbers from its launch post, and a reader should wait for the weights before treating any of them as independently verified.

Why the "sovereign" framing

Mistral is placing Large 4 between American closed models and Chinese open ones. VP Science Pierre Stock told TechCrunch the lab trained Large 4 with "two to three times less" GPU than its Chinese competitors and "significantly less than the closed source competitors", and said open-weight models are easier to audit than closed ones.

French president Emmanuel Macron has described this positioning as "a third way in AI", TechCrunch's Anna Heim notes. The company counts ASML, which led its Series C, and Samsung, which led its Series D at a reported €21 billion valuation, among its backers; both have reason to want a European option, and Mistral's VP of Science named cybersecurity, finance and chip design as the areas where Large 4 is tuned.

What a team actually gets

Right now a team can hit the preview on Mistral Studio and the API. The weights come at the end of October, which is the point at which a team that needs to self-host for compliance or privacy can start treating Large 4 as a real option. Until then the strongest claim, the cyber lead over refusing closed models, cannot be checked outside Mistral.

Source

Introducing Mistral Large 4, by Mistral AI, 6 October 2026. Reported by Anna Heim in TechCrunch, Mistral's new 1T model aims to leapfrog closed and open rivals.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.