Mistral Debuts “Le Chonk”: Open-Weight Large 4 in Public Preview, Weights Drop Oct 27

What Happened

The Wall Street Journal reported on October 6 (ET) that Mistral on Tuesday launched a public preview of the open-weight version of Large 4, codenamed “Le Chonk.” Unlike its usual direct releases, this is a two-step rollout: starting today, developers, cybersecurity experts and state authorities can evaluate the model's behavior via API; the weights — the trained parameters — won't be published until October 27, when anyone can download, self-host and fine-tune them. Mistral's official news page now lists “Introducing Mistral Large 4” at the top; when this column checked at 8:43 AM ET the entry wasn't there, so it went live within the past hour.

Hours before the launch, CEO Arthur Mensch teased it onstage at the Ai Everything conference in Abu Dhabi (first reported by Reuters). He said the model being announced later in the day is “above the Chinese models on certain aspects, including cyber,” adding that “the narrative that Europe cannot compete is something that is not true.” But Reuters' own copy notes the caveat plainly: he didn't specify which Chinese models, nor by what measure or with what scores. Mistral co-founder and chief scientist Guillaume Lample told the Journal: “The model capabilities will further improve as we scale up our training capacity.”

Key Facts

1. Preview today, weights Oct 27: Starting today it's API-only; the weights land October 27. Until then, no third party can independently reproduce or verify Mistral's performance claims — "among the top open-weight models globally" and "strongest outside China" are, for now, Mistral's own statements. The official line is that the preview lets the company "work with developers, cybersecurity experts and state authorities to evaluate how the model behaves."
2. ~4,000 Nvidia Grace Blackwell GPUs, trained in Europe: Mistral told the Journal the new model was trained on about 4,000 Nvidia Grace Blackwell GPUs deployed at its own data centers in Europe. VentureBeat separately reports the model at 1 trillion parameters — single source; neither Mistral nor the Journal has confirmed the parameter count.
3. "Closed models block cyberdefense requests": Mistral told the Journal that closed models "sometimes block responses to useful cyberdefense requests," which could pose security risks for users. Mensch's Abu Dhabi pitch was aimed at buyers who don't want to choose between American and Chinese infrastructure — governments and enterprises across the Gulf and APAC.
4. "Beats Chinese models": no numbers given: Per Reuters, Mensch named no Chinese model, cited no benchmark and gave no score. Mistral's official line is "among the top open-weight models globally on aggregate benchmark performance, and the strongest developed outside China by a substantial margin." The community can run its own evaluations once the weights are released on October 27.

Context

Mistral is Europe's leading AI lab: three years old, and smaller than OpenAI and Anthropic. In September it closed a €3 billion Series D (Reuters: €21 billion post-money; the Journal says the round was led by Samsung at a valuation above $24 billion), and its last model launch was in April. The company has consistently released open-weight models while earning revenue from its top model. This time it is opening a flagship-class model, as Chinese open models such as DeepSeek, Qwen and GLM keep shipping in 2026.