Mistral Unveils ML4 “Chonk,” a 1‑Trillion‑Parameter Open‑Weight Model

Overview

• Mistral released Mistral Large 4 (ML4), nicknamed “Chonk,” as an open‑weight model that can process text and images and is trained in over 160 languages.
• The model is available for preview on Mistral Studio now and the company has scheduled publication of the full weights for October 27 after planned audits and consultations with cybersecurity experts and authorities.
• Mistral claims ML4 matches top open‑weight performance and cites preliminary measures including a 38 score from independent site Artificial Analysis versus a market median of 26, though wider independent validation is still pending.
• The company says the model was trained on a comparatively modest compute footprint of roughly 4,000 NVIDIA Grace Blackwell GPUs and highlights strengths in cyber‑defense, industrial tasks, finance, and visual grounding.
• The launch follows a €3 billion funding round that valued Mistral at about €21 billion and is pitched as a European counterweight to U.S. and Chinese frontier models while raising questions about safety and real‑world robustness once weights are public.