Mistral’s new 1T model aims to leapfrog closed and open rivals

The European AI landscape is witnessing a fierce competition between open and closed models. On Tuesday, French AI lab Mistral AI unveiled their latest creation, Mistral Large 4 (ML4), a multimodal model aiming to surpass both American and Chinese counterparts. This ambitious move aligns with French President Macron's vision of an "AI third way."

Mistral's ML4, nicknamed "Le Chonk" due to its 1 trillion parameters, represents a significant leap in size, though it is still not an open-weight model. Currently, access to ML4 is restricted through a guardrail endpoint, pending safety testing. Mistral VP of Science, Pierre Stock, explained that open-weight models are more conducive to auditing for security purposes, a concern that has been on the rise among their primary clients—enterprises and institutions.

In a strategic move, Mistral AI has trained ML4 entirely on their own compute resources, utilizing only 4,000 NVIDIA GPUs, which is considerably less than their Chinese competitors and significantly fewer than closed-source rivals. While benchmark results are yet to be released, Mistral AI is confident that ML4 will be among the top open-weight models globally, particularly outside of China.

The company believes that ML4's optimized use cases, particularly in cybersecurity and finance, could also outperform closed models in specific areas of value to its customers. Notably, Mistral AI has chosen to forgo hosting Chinese models, positioning itself as a frontier lab despite the move.