Overview
- Mistral published the model weights on Thursday, October 29, 2026, making Mistral Large 4 (nicknamed “Le Chonk”) available for download on Hugging Face and through Mistral’s API.
- The company says three weeks of reinforcement learning raised the model’s Terminal‑Bench 4.0 score from 28% to 41%, and that ML4 narrowly tops other open models on agentic coding tests.
- Mistral reports ML4 is more efficient than a leading Chinese open model, using about 47% fewer tokens per task, and can run on modest hardware so organizations can deploy it on‑prem.
- The lab pitched ML4 for defensive cyber work such as automated pen testing and incident response, saying open weights protect business continuity because access cannot be cut off.
- Security experts caution that freely downloadable weights are dual‑use and could be repurposed by attackers, and journalists note proprietary models still lead on overall benchmarks such as Anthropic’s higher score.