Mistral releases 1-trillion-parameter open model le Chonk

Mistral unveiled its new flagship open AI model on Tuesday, describing it as the most capable open-weight system developed outside China.

Mistral Large 4 enters preview with targeted capabilities. The 1-trillion-parameter model, also called le Chonk, is positioned for tasks in cyber, coding, manufacturing, finance and multimodal applications. It is currently available in preview to developers, cybersecurity leaders and state authorities before a wider release later this month.

Training details and performance claims. ML4 was trained on 4,000 Nvidia Grace Blackwell GPUs over two months inside Mistral's European data centers. When its core parameters are released, the company states the model will rank among the top open-weight models globally on aggregate benchmarks and will lead all open-weight systems developed outside China by a substantial margin.

Limitations and future scaling plans. The model still trails frontier systems in coding performance. Co-founder and chief scientist Guillaume Lample said capabilities will improve as training capacity scales following the company's recent Series D round.

Context on open models and global competition. Open-weight models allow modification and self-hosting, unlike closed systems from OpenAI and Anthropic. Chinese open models have led in capability and adoption, prompting Western developers including Mistral to accelerate efforts. Earlier this year Hugging Face used a Chinese GLM model during a breach after U.S. closed systems blocked defensive actions, and Reflection AI released its own open model on Monday claiming parity with GLM 5.2.