A Trillion Parameters, Nicknamed “Le Chonk”: Europe Finally Has a Flagship
French AI company Mistral on October 6 unveiled its most powerful open-weight model to date: Mistral Large 4 (ML4), unofficially nicknamed “Le Chonk.” It is a 1-trillion-parameter, natively multimodal mixture-of-experts model with roughly 52 billion active parameters per inference, supporting more than 160 languages — every official language of the European Union included. “That one took some groundwork,” CEO Arthur Mensch posted on X, adding that the model was “trained and served on our own compute,” and that reinforcement learning “shows no sign of saturation.”
Le Chonk is Mistral’s first major launch since raising €3 billion in September — Europe’s largest-ever tech fundraise, valuing the company above €21 billion. A public preview API is live on Mistral Studio, with full weights promised by end of October after real-world red-teaming concludes. The story climbed to nearly 2,000 points on Hacker News, making it the AI community’s hottest topic of the week.
The Real Shock Is Cybersecurity: 82% vs. Zero
ML4’s most striking numbers are in cybersecurity. On one Artificial Analysis Cyber Index test — where a model must reproduce a real vulnerability in open-source software and then patch it — ML4 scored 82%, the highest of any model in the world. It also solved 93% of Cybench, a suite of 40 security-competition exercises. By contrast, Claude Opus 5.5 and GPT-6 Astra scored near zero on the same test because their safety filters refuse the task outright.
That contrast exposes the practical fault line between open and closed AI: defending software often begins with proving a flaw is real, exactly the kind of work that closed-model safety filters block. Chief scientist Guillaume Lample put it bluntly: “The cyber defence capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyberattacks.” With rogue-AI-agent incidents multiplying and attackers jailbreaking the very models enterprises rely on, the pitch lands hard with security teams.
3,800 Nvidia Chips and the Case for Sovereign AI
ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs, all inside Mistral’s own European data centers; the preview API is served on the same infrastructure. “The model we’re announcing today is above the Chinese models on certain aspects, including cyber,” Mensch told journalists. “The narrative that Europe cannot compete is something that is not true.”
Open weights let enterprises and governments download the model, customize it, and run it on their own servers — data never leaves the building. Mistral says clients in finance, manufacturing and the public sector can “govern the intelligence, data, compute and operations underpinning their critical infrastructure without exposing their most valuable knowledge to anyone outside their own walls.” The model will be available across multiple regions, including a European deployment that Mistral operates end-to-end under European law.
Three Takeaways for the Industry
First, the open-weight camp has a Western flagship again. For the past year, open-model leaderboards were dominated by Chinese releases; ML4 now tops the Artificial Analysis Intelligence Index among non-Chinese open models and even surpasses frontier closed models on some visual grounding benchmarks. The open race is no longer “China vs. everyone else” — it is a three-way contest.
Second, safety refusals have officially become a competitive liability. When closed flagships blank on practical security tests while attackers jailbreak those same models, the market will split faster between “models that can do the work” and “models that won’t” — and enterprise procurement may tilt toward controllable open deployments.
Third, sovereign AI has graduated from slogan to product. With owned compute, owned data centers and open weights, Mistral has packaged “your data stays yours, your policy is yours” into something you can order. Backed by Europe’s largest tech fundraise, the approach finally has the scale to negotiate with US and Chinese giants.
Conclusion: The Weights Release Is the Real Exam
Le Chonk’s preview benchmarks are impressive, but an open model’s true verdict always arrives after the weights ship — when independent deployment costs, fine-tuning difficulty and ecosystem uptake get tested. Around October 27, when a trillion parameters land on servers worldwide, we will find out whether “Europe cannot compete” was ever true. For enterprises, the preview API on Mistral Studio is worth trying now to gauge whether ML4 can take over your security and agent workflows.
