Mistral has launched a public preview of Mistral Large 4 — internally dubbed "le Chonk" — a 1 trillion-parameter mixture-of-experts model with 49 billion active parameters. The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. Weights will be released by end of month. This is the largest open-weight model developed outside the US-China axis. The headline capability claim is cybersecurity. ML4 scores 82% on a vulnerability reproduction-and-patch test from the Artificial Analysis Cyber Index — the highest of any model tested. It solves 93% of Cybench challenges. The pointed comparison: Claude Opus 5.5 and GPT-6 Astra score near zero on the same reproduction task because their safety filters refuse the work entirely. Mistral is making the explicit argument that closed-model refusals create a security gap for defenders. On coding and agentic benchmarks, ML4 posts 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and a combined Coding Agent Index of 49.8%, beating DeepSeek V4 Pro 0813 and Qwen3.8 Max. In blind human evaluation by Surge AI, it ranked second of five models at 3.74, behind Claude Opus 5 at 4.22. On AutomationBench (657 business workflows), it scores 59.9%. On financial and legal benchmarks evaluated by vals.ai, Mistral claims ML4 exceeds GPT-6 Astra. The multimodal story is dense visual grounding: satellite imagery, engineering drawings, document analysis. ML4 hits 42% on Dense 200, edging GPT-6 Astra's 41%. On science, it claims state-of-the-art among open-weight models on SciCode-Verified and can generate a complete Hartree-Fock simulation in one shot. Training data spans 160+ languages including every official EU language. The sovereignty architecture is the strategic core. Mistral trained, serves, and will deploy the model on its own European infrastructure, explicitly independent of other digital service providers and under European law. The red-teaming process involves cybersecurity leaders, vetted partners, and state authorities who receive reduced moderation and expanded cyber capabilities. This is Mistral positioning itself as the defense-grade AI provider for governments and enterprises that cannot or will not depend on US cloud platforms. The business model question is whether open weights at this capability level can sustain the compute investment. Mistral is signaling a forge-and-customize approach: the same training and RL environment used for ML4 is offered to enterprise customers through Mistral Forge. Specialized and optimized derivative models are promised. The revenue engine is customization and deployment services, not API margin. What matters structurally is the existence proof. A European company, on European hardware, in European datacenters, has produced a model competitive with or exceeding frontier closed models on specific enterprise verticals. Whether ML4 is the best model overall matters less than the fact that it demonstrably exists at this performance tier. The sovereignty argument stops being theoretical when you can point to a trillion parameters trained in Paris.
