Mistral has opened a public preview of Mistral Large 4, a natively multimodal model with 1 trillion total parameters and 49 billion active, informally named “le Chonk”. The company says ML4 outperforms every other open-weight model built in the US or Europe and is state of the art among open models on cybersecurity, finance and law, and claims it goes further still on visual grounding, beating frontier closed models such as GPT-6 Astra on the Dense 200 benchmark by 42% to 41%. It was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own European datacenters, and the preview API is served from that same infrastructure, but the open weights are promised only by the end of October. Until then Mistral is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who get access with reduced moderation and expanded cyber capabilities.

Related: Mistral Releases Medium 3.5 and Remote Agents in Vibe Environment, Aleph Alpha Open-Sources Kolibri 1, a Bilingual German-English MoE Model

Introducing Mistral Large 4 · Mistral AI announcement