Mistral Publishes ML4 Weights After Three Weeks of RL Tuning

Overview

• On Thursday, October 29, Mistral posted the Large 4 (ML4) model weights to Hugging Face and opened API access and selected Vibe users.
• Mistral says it ran three weeks of reinforcement‑learning fine‑tuning with customers and cybersecurity partners to improve performance.
• The company reports coding success on Terminal‑Bench 4.0 rose from 28% to 41% and that ML4 scores 79% on the CyberGym‑E2E security test, figures Mistral attributes to the RL tuning.
• Mistral claims ML4 needs about 47% fewer tokens per task than GLM‑5.3, implying lower inference cost, though some benchmark details have not yet been independently verified.
• Backed by roughly €3 billion in new funding, Mistral plans more frequent flagship releases and European data centers so companies can run ML4 on‑premises, reducing vendor lock‑in and easing compliance for sectors like health and finance.