Mistral Large 4 'Le Chonk': Europe's 1T AI Bet

Mistral Large 4, nicknamed "Le Chonk," is France's new 1-trillion-parameter open-weight AI model, trained in two months on 3,800 NVIDIA chips.

On October 6, 2026, French AI company Mistral launched the public preview of Mistral Large 4, or ML4 — its biggest and most capable model to date. You can try it today through an API on Mistral Studio, and the company says the model weights will be released by the end of October, so anyone can download and run it.

Meet Le Chonk: a trillion parameters, wide open

The headline number is staggering: 1 trillion parameters. A parameter is a tiny piece of knowledge the model learned during training — think of them as dials the model tunes to understand language, images, and logic. More parameters usually means a smarter model, but also a bigger, more expensive one.

Mistral got around that with a trick called a mixture of experts. Only 49 billion of the trillion parameters actually "fire" for any single answer — about 5 percent. It is like having a hospital with a thousand specialists on staff, but only calling in the few you need for your visit. The result is frontier-level smarts at a fraction of the computing cost.

The model is also natively multimodal, meaning it understands images directly, not just text. It can read documents, charts, technical drawings, and even massive satellite images, then zoom in and verify details. And crucially, it is open-weight: when the weights drop, you will be able to download the model itself and run it on your own servers — no data sent to anyone else.

Forged in Europe: 3,800 GPUs, two months

Mistral trained ML4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe — in roughly two months. CEO Arthur Mensch noted the training run finished so fast that the reinforcement learning had "no sign of saturation," meaning the model was still improving when they stopped.

The training data spans more than 160 languages, including every official language of the European Union. Mistral positions the whole project as a milestone for AI sovereignty: a state-of-the-art model built and served on European infrastructure, under European law, as an alternative to American closed models and Chinese open-weight systems.

How good is it? The benchmarks

Mistral says ML4 is competitive with the strongest open-source models in the world and "significantly outperforming any open-weight model developed in the US or Europe." Here are the headline numbers, straight from the company's announcement:

Image: Mistral

• Cybersecurity — 93% on Cybench. ML4 solves 93% of 40 challenges drawn from real security competitions, and scores 82% on an independent test asking it to reproduce a real vulnerability and then patch it — the highest score of any model. For more Cybersecurity news like this, see our label page.
• Coding — 61.7% on DeepSWE v1.1. In a blind human evaluation of coding quality, professional annotators ranked ML4 second out of five models — ahead of Kimi K3 and GLM-5.3, behind only Claude Opus 5.
• Visual grounding — beats closed frontier models. On the Dense 200 benchmark, ML4 scored 42% versus 41% for GPT-6 Astra, finding tiny objects in dense images better than a closed frontier model.
• Knowledge work — 59.9% on AutomationBench. The model completed real business workflows across Gmail, Google Sheets, Slack, and Salesforce, and independent evaluators found it beats GPT-6 Astra on legal and financial tasks.

The cybersecurity scores come with a sharp edge: Mistral notes that several leading closed models — including Claude Opus 5.5 and GPT-6 Astra — score near zero on the same vulnerability test because their safety filters refuse the task. For defenders doing legitimate security work, that refusal is itself a problem.

The sovereign AI pitch: your model, your rules

Customers will have two ways to use ML4: download the weights to their own servers, or access it through Mistral's infrastructure — including a European deployment where data stays under EU jurisdiction. Chief Scientist Guillaume Lample said the model's cyber defence capabilities will help enterprises and governments defend themselves against attackers who are jailbreaking closed models for attacks.

The timing is no accident. This is the first milestone on the roadmap funded by Mistral's €3 billion Series D — the largest equity round ever raised by a European technology company, valuing Mistral at over €21 billion. The money is already being spent on scaling compute in Mistral's own European datacenters, with more capacity coming online in the months ahead. For more AI news, browse our AI label page.

What's next, and what it means for you

Mistral is clear that this preview is just the start. The reinforcement learning run behind ML4 is still in flight, and more architecture details, benchmarks, and training methods are promised alongside the weights at the end of the month. Before the weights ship, the company is red-teaming the model with cybersecurity leaders, vetted partners, and state authorities.

Image: Mistral

• Developers: try the preview API on Mistral Studio today — especially if you build coding assistants or security tools.
• Security teams: this could be the first top-tier defensive AI you can run on your own infrastructure, with no provider able to pull the plug mid-incident.
• European companies and governments: a serious option for keeping sensitive AI workloads under EU jurisdiction.
• Everyone else: another heavyweight in the open-weight race means more competition, more choice, and eventually better and cheaper AI for all of us.

Mistral ended its announcement with a promise: "The pace of progress from here will be fast." With the weights landing at the end of October, the open-source community is about to find out just how fast.

Sources: Mistral — Introducing Mistral Large 4, Malay Mail (AFP), Chosun Biz