Mistral has released a public preview of its largest model to date — Mistral Large 4, internally nicknamed 'le Chonk.' It is a 1 trillion-parameter mixture-of-experts model with 49 billion active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. Weights are promised by end of month. The model is natively multimodal and targets enterprise verticals: cybersecurity, finance, law, and scientific computing. The cybersecurity angle is the sharpest differentiator. On the Artificial Analysis Cyber Index, ML4 ranks in the global top five and leads all open-weight models developed outside China. It scores 82% on a vulnerability reproduction-and-patch test — the highest of any model — and solves 93% of Cybench's 40 security challenges. Mistral frames this as a feature, not a risk: closed models like Claude Opus 5.5 and GPT-6 Astra score near zero on the same reproduction test because their safety filters refuse the task. Defenders, Mistral argues, need the same capabilities attackers are jailbreaking into existence. 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 score of 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In blind human evaluation by Surge AI, it ranked second of five models at 3.74 out of 5, behind only Claude Opus 5 at 4.22. On AutomationBench's 657 business workflows, it scores 59.9%. On knowledge-work deliverables (AA-Briefcase), it hits 1,393 Elo, again ahead of DeepSeek V4 Pro. The multimodal capabilities extend to visual grounding, where Mistral claims ML4 surpasses GPT-6-Astra on the Dense 200 benchmark (42% vs 41%). The model handles gigapixel satellite imagery, engineering drawings, and complex document reasoning. On scientific computing, it reaches state-of-the-art among open-weight models on SciCode-Verified and can generate a complete Hartree-Fock simulation in one shot. Third-party evaluators at vals.ai found it exceeds GPT-6-Astra on both legal and financial task benchmarks. The sovereignty framing is deliberate and structural. ML4 was trained on European infrastructure, served from European datacenters, operates under European law, and includes training data spanning 160+ languages including every official EU language. Mistral is positioning this as end-to-end independence from US cloud providers — not just a model, but an infrastructure play. The red-teaming phase before weight release involves cybersecurity leaders, vetted partners, and state authorities accessing the model with reduced moderation. The competitive landscape this enters is crowded. DeepSeek, Qwen, GLM, and Kimi are the named open-weight rivals; Claude, GPT-6 Astra are the closed-model benchmarks. Mistral's pitch is that ML4 matches or beats the best open models globally while being the strongest model developed in the US or Europe — a pointed acknowledgment that Chinese open-weight models (DeepSeek, Qwen, GLM) currently dominate the open frontier. The weight release transforms this from an API product into infrastructure that enterprises can self-deploy. What makes this more than a product launch is the policy argument embedded in the architecture. By shipping open weights with full cyber capabilities and no provider-level refusals, Mistral is arguing that safety-through-restriction is itself a security risk — that defenders lose when they cannot match attacker capabilities. This is a direct challenge to the alignment-via-refusal approach favored by Anthropic and OpenAI, and it will force regulators to take a position on whether open cyber-capable models are a net positive or negative for security.
