Overview
- Mistral has published the finalized weights for Mistral Large 4, nicknamed Le Chonk, on Hugging Face and says the model is now widely available for self‑hosting.
- The company cites Artificial Analysis benchmarks showing Le Chonk scoring about 41% on agentic coding tests versus GLM‑5.3’s 40% and 79% on a defensive cybersecurity index versus GLM‑5.3’s 29%.
- Mistral says the model improved after reinforcement learning during a short preview period and was trained on roughly 3,800 Nvidia Grace Blackwell GPUs using about 10 megawatts, which the firm presents as evidence of compute efficiency.
- Security experts warn the public release is a dual‑use risk because the same downloadable model that helps automate penetration testing can be adapted by attackers to find or exploit vulnerabilities.
- Mistral plans a family of Le Chonk models, including smaller laptop‑capable versions, but independent benchmarking and oversight are still needed to corroborate the company’s performance and safety claims.