Mistral holds back Large 4 weights for about three weeks of safety testing, citing cyber risk

French lab Mistral AI released Mistral Large 4, a multimodal model nicknamed Le Chonk for its roughly one trillion parameters, TechCrunch reports (technical details). It is positioned as an alternative to both American closed models and the open models increasingly coming from China, the “third way” French President Emmanuel Macron has described. A separate account of its industrial focus is in our earlier piece (Le Chonk’s niche bet).

The rollout is the notable choice. For now the model is reachable only through a guardrailed public endpoint. Mistral says it plans to release the weights in about three weeks after safety testing, and to work with trusted partners and governments in the meantime so the weights help defenders more than attackers.

Mistral says the model was trained entirely on its own compute using 4,000 Nvidia GPUs, which VP of science Pierre Stock says is two to three times fewer than Chinese competitors use and far fewer than closed labs. Training a model this size on a few thousand GPUs would suggest the frontier is not purely a function of who owns the most chips, a theme also seen in Alibaba’s scaling path (Qwen’s story).

Benchmark results were still pending at publication, and leapfrogging rivals is a stated aim rather than a demonstrated result. Until independent numbers arrive, the supportable claim is that Mistral has built a large, targeted model. Mistral recently began hosting Chinese models on its platform and has worked to show that was not a pivot to being only an inference provider; Large 4 is its case that it remains a frontier lab.

Mistral says the model is tuned for cybersecurity, finance and chip design. Chip design is central to ASML, which led Mistral’s Series C, and Samsung, which led its Series D the previous month at a €21 billion valuation. Building for backers’ industries is one way a smaller lab can beat larger general-purpose models on specific work.

Stock argues that open weights raise misuse risk but are far easier to audit than a model reachable only through an API. A trillion-parameter, cyber-tuned open model is useful to defenders and attackers alike, as Mistral acknowledges. Three weeks of testing and a partner program are a start rather than a guarantee; if the pattern spreads, the industry may need shared norms for what a pre-release window must demonstrate. More in our cybersecurity coverage, research coverage and Nvidia coverage.