One Trillion Parameters, 49B Active: Mistral's Le Chonk Is Europe's Boldest Open-Weight Bet Yet

On October 6, 2026, Mistral AI launched the public preview of Mistral Large 4 — unofficially ML4, and, in the company’s own words, “very officially: le Chonk.” It is the Paris-based lab’s largest and most capable model to date, and it arrives carrying an unusual pair of claims: frontier-class performance in specific enterprise verticals, delivered through open weights that will be published by the end of October.

The numbers are the headline. ML4 is a 1-trillion-parameter, natively multimodal mixture-of-experts model with 49 billion active parameters per token. Mistral trained it from scratch over roughly two months on 3,800 NVIDIA Grace Blackwell GPUs running in its own European data centers — infrastructure the company built after raising €830 million in debt financing earlier this year for a GB300 fleet and a 44-megawatt facility near Paris. The public preview is served on that same infrastructure, and Mistral says the model will be available across multiple regions worldwide, including a European deployment it operates end-to-end, independently of other digital service providers and under European law.

What the benchmarks actually say

Mistral’s announcement is unusually specific for a preview launch, and the cybersecurity results are the most striking. On the Artificial Analysis Cyber Index — an independent evaluation of how well models find and fix security flaws in real software — ML4 ranks among the top five models globally and leads open-weight models developed outside China by a wide margin. On one of the index’s tests, which asks a model to reproduce a real vulnerability in open-source software and then patch it, ML4 scores 82%, the highest of any model tested. It also solves 93% of Cybench, a set of 40 exercises drawn from security competitions — one of the highest scores ever reported for an open-weight model.

The comparison Mistral draws is pointed: Claude Opus 5.5 and GPT-6 Astra reportedly score near zero on the same vulnerability-reproduction test, not because they lack the capability, but because their safety filters refuse the task. Mistral frames this as the practical case for open weights in security work — defending software often starts with proving that a flaw is real, and a model that declines to do so is of limited use mid-incident. Red-teaming for the preview period is being run with “reduced moderation and expanded cyber capabilities” for vetted partners, cybersecurity leaders, and state authorities.

On coding more broadly, ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, for a combined Coding Agent Index of 49.8% — ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. A blind human evaluation run with Surge AI, in which professional annotators rated outputs on a 1–5 scale with model identities hidden, placed ML4 Preview second of five models at 3.74, ahead of Kimi K3 (3.59), GLM-5.3 (3.60), and GLM-5.2 (3.40), behind only Claude Opus 5 (4.22).

Agentic results follow the same pattern. On AutomationBench — 657 business workflows across Gmail, Google Sheets, Slack, and Salesforce — ML4 scores 59.9%, ahead of Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro. And in visual grounding, Mistral claims the model surpasses even frontier closed models, going beyond what open weights had previously managed.

Why the “sovereign AI” framing matters

Strip away the benchmark table and the launch is really an argument about deployment models. Mistral positions ML4 as “forged in Europe, built for AI sovereignty” — a model that enterprises and governments can self-host, audit, and run under their own policies, on-premise or in private cloud. That pitch lands hardest in cybersecurity (where provider-level refusals can block legitimate vulnerability research), in regulated sectors like finance and law (where Mistral claims state-of-the-art results among open models), and in the growing number of jurisdictions demanding data and compute residency.

The training story reinforces it. A significant share of ML4’s training data was multilingual, spanning more than 160 languages, including every official EU language. Mistral also notes the model was trained on the same customization and RL environment it offers customers through Mistral Forge — meaning enterprise fine-tuning and the base training pipeline share a toolchain.

The open-weights race has a new frontrunner outside China

For most of 2026, the open-weight frontier has been a Chinese affair — DeepSeek, Qwen, Kimi, and GLM trading the crown back and forth while Western labs kept their strongest models closed. Mistral now claims ML4 “significantly outperform[s] any open-weight model developed in the US or Europe,” and while weights won’t land until October 27, the preview benchmarks back the claim well enough that Wired described the model as by far the most capable open-weight model developed outside China, and “very, very close” to some proprietary frontier models.

There are caveats worth holding onto. These are Mistral’s own submitted numbers on several benchmarks; independent verification at scale only becomes possible when the weights drop. A 1T-parameter MoE is expensive to self-host even at 49B active — the sovereignty pitch assumes customers who can operate Grace Blackwell-class infrastructure or rent it. And “preview” means the model is still being refined; Mistral itself says it “continues to improve rapidly.”

But the strategic read is clear. With a €3 billion Series D led by Samsung behind it (a post-money valuation above €21 billion, the largest equity round in European tech history), acquisitions stacking up across industrial and creative AI, and now a 1T-parameter flagship with a self-deployment story no American hyperscaler can copy, Mistral is no longer the plucky open-source alternative. It is making the case that the next open-weight frontier model can be European — and it is putting a trillion parameters behind that case.

The weights are due by the end of October. Until then, the preview API is live on Mistral Studio, and the industry gets three weeks to argue about whether a model named after an internet meme can carry a continent’s AI ambitions.