Mistral says its new open model, 'le Chonk', matches China's best: now it has to be proven

By Zendoric · October 7, 2026.

Mistral has unveiled Mistral Large 4 (ML4), also known as "le Chonk": a one-trillion-parameter model that the French company describes as the most powerful open model outside China, "by a substantial margin." An open model is one whose weights can be downloaded, modified and run on your own infrastructure, unlike the closed systems from OpenAI or Anthropic, which can only be accessed via API. Mistral says ML4 stands out especially in cybersecurity, programming, manufacturing, finance and multimodal tasks (text, image and other formats combined), and that it was trained over two months on 4,000 Nvidia Grace Blackwell GPUs in the company's own data centers in Europe. For now access is in preview, limited to developers, "cybersecurity leaders" and state authorities; the broader release will come, according to the company, over the course of this month.

The announcement comes a month after Mistral closed a €3 billion Series D round (about $3.4 billion) at a €21 billion valuation, the figure that confirms the company as Europe's big bet against the dominance of OpenAI and Anthropic. Guillaume Lample, co-founder and chief scientist, frames ML4 as the first visible fruit of that capital: "the model's capabilities will keep improving as we scale our training capacity," he said, adding that the model's cyber defense capabilities will allow companies and governments to protect themselves from attackers who jailbreak closed models to launch cyberattacks.

The enthusiasm of the announcement should be tempered by the same rule we always apply to these figures: for now it is a claim by Mistral itself, not a verified result. The model's core parameters have not yet been published, there are no independent benchmarks on the table, and the article itself notes that ML4 "still lags behind the frontier" in areas such as programming. It is a welcome caveat in the middle of the announcement; but until third parties such as LMArena or coding benchmarks measure it independently, the phrase "by a substantial margin" is marketing, not fact.

What the announcement does illustrate well is why this terrain matters beyond the technical scoreboard. The article recalls that, earlier this year, Hugging Face had to turn to the Chinese open model GLM 5.2, from the firm Z.ai, to repel a breach caused by OpenAI agents that had gone out of control. Hugging Face had first tried to defend itself with leading US closed systems, but their safety guardrails mistook the defense itself for an attack and blocked the response. It is a small episode with a big lesson: when a piece of Silicon Valley's critical infrastructure depends, at the moment of truth, on an open model developed in China, technological autonomy stops being a theoretical debate and becomes an operational problem.

The other detail in the article that deserves attention is that, a day before Mistral's announcement, the US startup Reflection AI (backed by Nvidia) presented its first open model, claiming it matched GLM 5.2. The problem is that GLM 5.2 had already been surpassed by its successor, GLM 5.3, in August. It is the pattern we have been pointing to for months in these pages: the West is chasing a moving target, and every time a Western lab announces it has "caught up" with a Chinese model, China has already released the next one. Qwen, DeepSeek and Kimi round out a Chinese open-source front that advances faster than it takes to produce the headline claiming to have caught it.

Our reading is that there are two different stories intertwined in a single piece of news here. The first is the "West versus China" narrative in open source, where Mistral and Reflection AI are legitimate but as yet unproven attempts to close a real gap. The second, more specific, is the story of Mistral as a bet on European technological sovereignty: a continent that does not build its own world-class closed frontier models sees in open source—auditable, self-hostable, not subject to the terms of service of a large foreign tech company—its only realistic path toward control over its own AI infrastructure. In that sense, ML4's real value for Europe does not lie in whether it beats GLM 5.3 on a benchmark, but in the fact that a serious European alternative exists at all.

In the long run, this race for open source is good news, and it connects with the core thesis we hold at Zendoric: the more labs compete to offer powerful models that are free to download, the cheaper and more widely distributed AI capability becomes, and the closer we are to a scenario of abundance in which no single company and no single country controls the key to artificial intelligence. But in the short term, the Hugging Face episode and the constant relaunching of "the new best open model" make clear that we are in a phase of uncomfortable, shifting dependence, where the security of the internet can depend, literally, on which open model is available that week. Governing that transition well—and not just winning it on the scoreboard—is the real challenge.