Mistral AI has released the weights for Mistral Large 4, its 1-trillion-parameter multimodal model nicknamed "Le Chonk," after a public preview period that the company says exceeded expectations. The release positions Mistral at the top of open-weight coding benchmarks and marks the company's clearest bet yet that open weights, European training infrastructure, and cybersecurity specialization constitute a viable competitive strategy against closed-model incumbents. The coding numbers are real and specific. On Cognition's private FrontierCode 2.0 benchmark, Mistral Large 4 leads with 19.7%, ahead of DeepSeek V4.1 Flash at 19.3% and GLM-5.3 at 17.8%. On the Artificial Analysis Coding Agent Index, it scores 48%, effectively tied with GLM-5.3 but using 47% fewer tokens per task. Terminal-Bench 4.0 performance jumped 13 percentage points from preview to 41%, with 30% fewer output tokens. Three reasoning modes — low, medium, and high — let developers trade precision for cost. The cybersecurity angle is where Mistral is trying to carve genuinely new territory. On the Artificial Analysis Cyber Index, Le Chonk scores 51%, measuring vulnerability discovery, reproduction, and patching. Crucially, the model achieves a 96.6% refusal rate on malicious cyber requests — the highest Mistral measures among open models — while still performing defensive tasks without unnecessary refusals. The company acknowledges observing "potential for uplift to conduct real attacks" in live traffic and says it is modifying guardrails accordingly. The enterprise pitch extends beyond security. FinWorkBench scores climbed to 73.3% for finance workflows, BrowseComp reached 84.5% for web browsing, and the B3 Agent Security Benchmark shows 97.1% attack resistance, up from 93.3% in preview. La Banque Postale is cited as a strategic partner, framing the relationship explicitly around "technological sovereignty" — a European bank choosing a European model. The sovereignty narrative is doing real work here. Mistral emphasizes the model was "trained from scratch on European infrastructure with clear lineage from inception to release." The company is positioning open weights not as charity but as a business model: organizations can self-host, customize via Mistral Forge, and distill task-specific variants. A specialized cybersecurity variant is being offered to enterprise partners like Orange Cyberdefense under "rigorous vetting and contractual safeguards." The competitive landscape is tight. Surge Code's blind human evaluations show Mistral Large 4 preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, but Claude Opus 5.5 still wins 78% of head-to-head matchups. The honest read: Le Chonk is the best open-weight coding model available right now, competitive with closed models on many tasks, but not yet the best model period. The token efficiency advantage — doing equivalent work with roughly half the tokens — may matter more to enterprise buyers than raw benchmark leads. Three weeks of additional reinforcement learning drove most of these gains from preview to release, with Mistral claiming "no signs of saturation." If that trajectory holds, the question becomes whether open-weight models at this scale can sustain improvement velocity against the compute budgets of Anthropic, Google, and OpenAI. Mistral is betting that for enterprise buyers who need to own their stack, "close enough and controllable" beats "slightly better but rented."
