Mistral Is SO Back: Inside Le Chonk, the West’s Biggest Open Model (Pierre Stock, VP Science, Mistral)

Mistral’s research lead Pierre Stock details the development of "Le Chonk" (Mistral Large 4), a trillion-parameter mixture-of-experts model trained on 3,800 GPUs in Europe. This release underscores the strategic necessity of open-weight models for ensuring sovereign AI access and business continuity amid geopolitical restrictions on closed-source technology. Training such massive models with limited compute relies on rigorous data calibration and sophisticated reinforcement learning techniques, which allow the model to achieve frontier performance while maintaining operational efficiency. By prioritizing open-weight releases, Mistral decentralizes AI power and fosters a collaborative research ecosystem. Future iterations aim to advance agentic capabilities, enabling AI to autonomously navigate complex security incidents and engineering tasks without relying on pre-written playbooks, thereby pushing the boundaries of what models can achieve through self-directed reasoning and real-time problem solving.