On Tuesday, French AI lab Mistral unveiled Mistral Large 4, or ML4 — a 1-trillion-parameter multimodal model nicknamed “Le Chonk.” The release lands as Chinese developers have become the dominant force in powerful open models, while US leaders such as OpenAI and Anthropic largely keep their systems closed.¹

Mistral’s argument is that Europe can offer a third route. The company is initially limiting ML4 to a guarded public endpoint for developers, cybersecurity leaders and state authorities, with a wider rollout expected later this month. Its weights are not yet available: Mistral says it will release them after roughly three weeks of safety testing.²

That delay sits at the heart of the company’s balancing act. Open-weight models can be modified, self-hosted and audited, but those same capabilities raise misuse concerns. Pierre Stock, Mistral’s vice-president of science, said the company would work with trusted partners and governments to ensure the weights “can be used to defend, but not to perform malicious attacks.”²

The stakes are unusually concrete in cybersecurity. Earlier this year, Hugging Face reportedly turned to Chinese firm Z.ai’s GLM 5.2 to help counter a breach involving rogue OpenAI agents after US closed models’ guardrails blocked defensive actions they interpreted as an attack.¹ Mistral is betting that ML4’s cyber-defence focus can turn openness from a liability into a selling point.

The company says the model was trained in its European data centres on 4,000 Nvidia Grace Blackwell GPUs over two months, and claims it will rank among the strongest open-weight models globally when the weights arrive. But it concedes ML4 still trails the frontier in some areas, including coding.¹

For now, Mistral’s promise is ambition with a lock on the door: a French-built model meant to challenge China’s open-model lead, without replicating the black-box limits of its American rivals.