Overview
- Mistral published the full weights for Mistral Large 4, nicknamed Le Chonk, on Oct. 29 and made the model available on Hugging Face plus a commercial API for teams that do not self‑host.
- The company says ML4 is a Mixture‑of‑Experts model with about 1 trillion total parameters, roughly 52 billion active per token, support for very long contexts, and the ability to run on as few as four Nvidia B200/B300 GPUs.
- Mistral reports reinforcement‑learning tuning raised ML4’s Terminal‑Bench 4.0 score from 28% in the preview to 41% and says the model matches or narrowly beats China’s GLM‑5.3 on many coding and cybersecurity tests while using about 47% fewer tokens per task.
- Mistral published safety metrics and an Apache‑compatible open‑use license that permits commercial use, claiming a 96.6% refusal rate on malicious cyber prompts, but observers note the license’s ban on “malicious use” is vague and external auditing and sustained testing remain outstanding.
- Microsoft will offer ML4 through Microsoft Foundry with options for regional deployment such as Sweden Central and built‑in governance tools, giving regulated customers a hosted path while raising questions about the extra infrastructure and operational work required to run and secure an open‑weight model.