Mistral AI has released 'Le Chonk,' a 1-trillion-parameter model, asserting its dominance in coding and cybersecurity among open-weight artificial intelligence systems. The Paris-based firm claims this new model, officially designated Mistral Large 4 (ML4), surpasses leading Chinese competitors, including those from DeepSeek and Z.ai, in crucial performance benchmarks. This development marks a significant claim in the ongoing competition for advanced AI capabilities, particularly in the open-weight sector.
ML4 was trained from scratch on Mistral's European infrastructure, utilizing 3,800 NVIDIA Grace Blackwell GPUs over approximately two months. The model is multimodal, capable of processing both text and images, and was trained on over 160 languages. Mistral has made a public preview of ML4 available, with plans to release the model weights under a custom license by the end of October 2026.
The company's announcement highlights ML4's performance in cybersecurity, stating it achieved a 93% score on Cybench and an 82% score on a reproduce-and-patch test, where it claims some leading closed-source models refuse to engage. Mistral also asserts that ML4 ranks among the top five models globally in cybersecurity and leads open-weight models developed outside China by a significant margin. Independent evaluations on the Artificial Analysis Cyber Index show ML4 tying for first place with Z.ai's GLM 5.3 Flash.
In coding benchmarks, Mistral claims ML4 leads on FrontierCode and Terminal-Bench, and matches the best open-weight models on the Artificial Analysis Coding Agent Index with greater token efficiency. A blind human evaluation commissioned by Mistral from Surge AI placed ML4 second out of five models in coding output quality, scoring 3.74 out of 5, behind Claude Opus 5 but ahead of models like Z.ai's GLM-5.3. However, some independent analyses suggest that while ML4 performs well, it trails other leading models, including some Chinese open-weight systems, on broader coding benchmarks like DeepSWE v1.1.
Mistral AI's assertion of superiority in coding and cybersecurity directly challenges the recent dominance of Chinese AI labs in the open-weight model space. Companies like DeepSeek and Z.ai have released models that have garnered significant attention for their performance and accessibility. DeepSeek, for instance, has released various models, including the V4 series, with competitive benchmarks in coding and general reasoning. Z.ai, formerly Zhipu AI, has also been a prominent player with its GLM series, known for its performance and integration with various hardware platforms. The release of Le Chonk positions Mistral as a key European contender aiming to balance advanced AI development with an open-weight philosophy.