Mistral Large 4 'Le Chonk' targets China's open AI models

Mistral Large 4, released October 6, 2026, arrives as an unusual open-weights push from Europe. Its 1.05兆 total parameters are split across a Mixture-of-Experts architecture, so only 490億 run per answer — a design that keeps the model deployable despite its size. The model is trained and served on Mistral's own infrastructure, which Mistral frames as a matter of "AI sovereignty" for companies and governments that do not want to depend on outside AI providers.

Mistral is explicitly pitching the model against Chinese open-weights rivals, pointing to a 49.8% Coding Agent Index and wins over DeepSeek and Qwen models on that measure. A separate blind grading test with Surge AI placed it second of five, behind Claude Opus 5 but ahead of GLM-5.3, Kimi K3 and GLM-5.2. Mistral also claims state-of-the-art results on cyber defense, finance and legal workloads, and says it slightly beat GPT-6 Astra on visual grounding.

The near-term test is the promised weights release at the end of October, when Mistral says it will publish architecture details and additional benchmarks. Until then, the model is only a Public Preview via API, and much of the evidence rests on Mistral's own evaluations. Whether the Le Chonk bet pays off may hinge on how outside developers respond once they can run it themselves.