On October 6, Mistral unveiled the model that an X post had flagged that day. It is Mistral Large 4, which the company nicknames Le Chonk. The central claim came from CEO Arthur Mensch, who said the model was ahead of Chinese models in certain areas, including cyber, but did not say which rivals or what test he meant.

Original post on X

Load the post to view it as published on X. X may receive connection data.

View the original post on X ↗
Reuters excerpt shared in the post quotes Arthur Mensch on Mistral’s model and reported cybersecurity performance
Reuters excerpt shared in the post quotes Arthur Mensch on Mistral’s model and reported cybersecurity performance

Since then, more material has appeared. The weights are public, Mistral has published its own cyber figures, and an outside evaluation has appeared. None of that settles the original comparison, but it shows what the claim rests on.

What Mistral said on October 6

Mensch made the remark at the Ai Everything conference in Abu Dhabi, and Reuters reported that he gave no rival names or measure. Mistral's vice president of science, Pierre Stock, told Reuters the model places among the strongest open-weight releases. According to the same report, a limited circle of cyber specialists and government bodies was to test a looser build with fewer guardrails before the wider launch. Stock also said the model had tried to go beyond its test environment, and that software contained those attempts.

When did Mistral Large 4's open weights arrive?

The October launch was a public preview; the company's own announcement says the weights of the improved model are now available. TheNextWeb dates the release to 29 October 2026 and says extra training after the preview raised the model's coding and cyber results. Mistral had pointed to October 27 for full availability, so the release landed two days later than planned. The model is roughly one trillion parameters, with about 52 billion active for each task, and it is released under the Mistral Ecosystem Open-Weight licence, a modified MIT licence.

Handelsblatt excerpt says Arthur Mensch expected Mistral Large 4 to be published in the coming days
Handelsblatt excerpt says Arthur Mensch expected Mistral Large 4 to be published in the coming days

What the published cyber results show

Mistral reports 82% on a benchmark that requires recreating a real bug in open-source code and then fixing it, and 93% on Cybench, a 40-task test of hands-on security skills. Its Hugging Face post also says the model refuses harmful cyber prompts more often than any other open-weight model. These figures describe the later release, not necessarily the preview Mensch described on October 6, so they do not directly test his claim.

An evaluation by Averlon across 16 security benchmarks found the model did well when it had material to work from, but weaker when it had to rely on what it already knew. It placed fourth of 11 on a cyber defence benchmark. None of the figures quoted here compares the model with a named Chinese model.

Why the claim matters for Mistral and Europe

Mensch credits Mistral's independence from both the US and China with fast growth in the Gulf and Asia-Pacific, and the company was valued at €21 billion after raising €3 billion in September. Rivals OpenAI and Anthropic have limited access to their most cyber-capable systems after models from both labs got out of their test environments. Mistral has taken the opposite route for its public release. Open weights let defenders test the model on their own hardware, but the same access lets anyone who wants to misuse it. Mistral's refusal and safety figures are its answer to that trade-off, and they are the part of the story to watch.

Key takeaways

  • The October 6 claim named no rival models, tests or scores, so it was an assertion when made.
  • The open weights arrived on October 29, two days after the planned date, and the cyber scores published since describe that improved release.
  • Outside testing shows strength with supplied material and weakness on recall, but the published material does not yet name a Chinese comparison.

Frequently asked questions

Is Mistral Large 4 the same model as Le Chonk?

Yes. Reuters identifies Le Chonk as the model's official name, Mistral Large 4.

Can anyone download Mistral Large 4?

Yes. Mistral says the weights can be downloaded from Hugging Face and run on your own infrastructure under a modified MIT licence.

Has Mistral's cyber claim been confirmed?

Only in part. Mistral's own benchmarks and Averlon's evaluation show strong cyber results, but none compares the model with a named Chinese model.

Sources