French AI lab Mistral AI has unveiled its new model, ML4, a large multimodal model with 1 trillion parameters, positioning itself as a potential leapfrog alternative to both American and Chinese rivals. The model, nicknamed Le Chonk, is currently restricted to a public guardrail endpoint but will be made fully open-source in three weeks, subject to safety testing.

Mistral VP Science, Pierre Stock, has assured that the open-source weights will be used defensively, not offensively. This move comes amid rising security concerns among enterprise and institutional users, who are increasingly wary of open-weight models made in China.

The model was trained entirely on Mistral’s compute, using 4,000 Nvidia GPUs, which is notably less than the competition. This, combined with focused training, could result in ML4 outperforming closed models in specific areas crucial to its customers, such as cybersecurity and finance.

Mistral’s backers, including ASML and Samsung, are key to its success. With a €21 billion valuation, Samsung led the Series D last month, while ASML led the Series C. This positioning could help Mistral maintain its status as a frontier lab, despite the challenges posed by the global AI landscape.