On Tuesday, French AI lab Mistral unveiled Mistral Large 4, or ML4—nicknamed “Le Chonk”—and immediately framed its trillion-parameter model as more than a product launch. It is a bid to keep Europe relevant as Chinese developers dominate the open-model market and U.S. leaders keep their strongest systems closed.¹

Mistral says ML4 is particularly suited to cybersecurity, coding, finance, manufacturing and multimodal work. The company trained it in European data centres on 4,000 Nvidia Grace Blackwell GPUs over two months, and argues that, once its core parameters are released, it will stand among the world’s leading open-weight models. Its own assessment is bolder still: ML4 is the strongest open-weight model built outside China “by a substantial margin.”¹

For now, however, Le Chonk is not actually open-weight. It is available in preview through a guarded public endpoint, with its weights slated to arrive at the end of October.² That gap is central to Mistral’s argument. The company wants the auditability and deployability of open weights without creating an easy route for abuse.

Pierre Stock, Mistral’s vice-president of science, said the interim period would be spent with “trusted partners and governments” ensuring the weights could “be used to defend, but not to perform malicious attacks.”³ The company contends that stronger cyber-defence tools matter precisely because attackers are already trying to jailbreak closed models.

The ambition comes with caveats. Mistral acknowledges that ML4 still trails the frontier in some areas, including coding, while benchmark results remain pending. Co-founder Guillaume Lample nevertheless said capacity would improve as the company scales training after its €3 billion Series D: “The model capabilities will further improve as we scale up our training capacity.”¹

That leaves Mistral trying to prove two things at once: that Europe can build a credible alternative to American and Chinese AI, and that openness need not mean recklessness.