Mistral launches public preview of 1-trillion-parameter Mistral Large 4, weights due this month

A trillion-parameter preview

Mistral has opened a public preview of Mistral Large 4, unofficially nicknamed "le Chonk". According to the company's announcement, which reached the front page of Hacker News, the preview API is live on Mistral Studio and the model weights will be released by the end of the month.

ML4 is a natively multimodal model with 1 trillion parameters, of which 49 billion are active. Mistral describes it as its largest and most capable model to date, and says it continues to improve as the company refines it ahead of the weight release. The company claims it is competitive with the strongest open models globally and significantly outperforms any open-weight model developed in the US or Europe, with state-of-the-art results among open models on enterprise workloads in cybersecurity, finance and law.

Cybersecurity as a headline feature

The most striking part of the announcement is how much of it concerns security work. Until the weights ship, Mistral says it is red-teaming the model in real-world settings with cybersecurity leaders, vetted partners and state authorities, who get access to a version with reduced moderation and expanded cyber capabilities.

Mistral cites the Artificial Analysis Cyber Index, an independent evaluation of how well models find and fix security flaws in real software, where ML4 ranks among the top five globally and leads open-weight models developed outside China by a wide margin. On one index test that asks a model to reproduce a real vulnerability in open-source software and then patch it, the company reports a score of 82%, which it says is the highest of any model. It also reports 93% on Cybench, a set of 40 exercises drawn from security competitions.

Mistral contrasts this directly with leading closed models, claiming Claude Opus 5.5 and GPT-6 Astra score near zero on the same vulnerability test because they refuse the task. The company's argument is that defending software often starts with proving a flaw is real, and that safety filters in closed models can block exactly that work while attackers jailbreak the same models for offensive purposes.

Benchmarks and blind evaluations

On agentic coding, Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, with a combined Coding Agent Index score of 49.8% that it places ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.

The company also commissioned a blind human evaluation from Surge AI, in which professional annotators rated outputs on a one-to-five scale without knowing which model produced them. ML4 Preview ranked second of five models with 3.74, ahead of Kimi K3 (3.59), GLM-5.3 (3.60) and GLM-5.2 (3.40), and behind only Claude Opus 5 (4.22).

Other reported results include 59.9% on AutomationBench, a suite of 657 business workflows across apps such as Gmail, Slack and Salesforce; 1,393 Elo on the AA-Briefcase long-horizon knowledge-work benchmark; and a visual grounding result on Dense 200 of 42%, narrowly ahead of GPT-6 Astra at 41%. Mistral says ML4 is state of the art among open-weight models on SciCode-Verified and can generate a full Hartree-Fock chemistry simulation in one shot. All of these figures come from Mistral itself unless otherwise noted.

Trained in Europe, pitched on sovereignty

ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the public preview is served from that same infrastructure. Mistral says a significant share of the training data was multilingual, spanning more than 160 languages, including every official language of the European Union.

The model will be available across multiple regions, including a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law. Mistral also says ML4 was built with the same training, customization and reinforcement-learning environment it sells to enterprises through Mistral Forge, and that it will serve as the foundation for a new generation of specialized models. Further details on the architecture, additional benchmarks and post-training methodology are promised before the weight release.

Why it matters

Frontier-scale open-weight models are rare, and a 1-trillion-parameter release with weights promised within weeks would let organizations self-host a top-tier model under their own policies, which is the core of Mistral's sovereignty pitch. The cybersecurity emphasis cuts both ways: fewer refusals genuinely help defenders and incident responders, but the same capabilities raise obvious dual-use questions that the pre-release red-teaming is meant to probe. Until independent evaluations arrive, the benchmark numbers remain Mistral's own claims. The end-of-month weight drop is what will make them checkable by anyone.