The French company Mistral AI newly revealed Mistral Large 4, its new artificial intelligence model that has the nickname “El Chonk”. With 1,000 billion parameters, of which 49 billion active during each processing, it targets in particular the best Chinese open-weight models, while strengthening Mistral’s position in the race for cutting-edge AI models.
Training in Europe for the big 4 (Le Chonk)
Mistral Large 4 is based on a multimodal Mix of Experts (MoE) architecture and can process text and images. Mistral says it trained it completely from scratch for about two months with 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, while its training covers more than 160 languages, including all official languages of the European Union.
The AI model also has a million-token context window, allowing it to work with very large volumes of information in a single request. Mistral presents it as a general model capable of combining reasoning, programming, image understanding and agency functioning, with a particular positioning in cybersecurity, finance, law and industry.
Mistral boasts leading performance among open weight models. In the artificial analysis Cyber Index, Large 4 ranks among the top five global models and achieved 82% in a test aimed at reproducing and then fixing a real software vulnerability. It also scored 93% on Cybench, a set of 40 cybersecurity competition exercises.
Programming is also among the announced strengths. Large 4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, while Mistral scores 49.8% on its encoder agent index. In a blinded human review by Surge AI, the Preview version took second place out of five models tested, behind the Claude Opus 5.
Mistral wants to reduce the gap with Chinese models
The launch comes at a time when Chinese open weight models are taking center stage. Arthur Mensch, head of Mistral, and his teams present the Large 4 as a response to this progression with the ambition of offering the most efficient open-weight model developed in Europe or the United States. In particular, Mistral considers its new model to be superior to certain Chinese competitors in areas such as cybersecurity, although performance has yet to be confirmed by broader independent evaluations.
However, the model is not limited to code and cybersecurity. Mistral also claims to have cutting-edge performance in finance, manufacturing, scientific analysis and image understanding, with “visual connection” capabilities that would even outperform certain proprietary AI models in the company’s internal benchmarks. Mistral claims, for example, to score 42% versus 41% for OpenAI’s GPT-6 Astra in Dense 200.
Large 4 is available today in public preview via the Mistral Studio API. However, the company won’t release model weights until October 27. Until then, Mistral will test the model in real conditions with cybersecurity specialists, selected partners and public authorities.
Meet Mistral Large 4, also known as The Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best US or European open weighting model in aggregate benchmarks.
• Latest in critical workloads, including cyber defense, manufacturing and finance, and outperforms closed-frontier models in terms of visual foundation.
• Forged in Europe from end to end and can be deployed from Europe through our own Mistral Cloud infrastructure.
• Available to everyone via API today. Work with cybersecurity partners privately.Open weight drop at the end of October.
—Mistral AI (@MistralAI) October 6, 2026
This testing phase is particularly important for a model with advanced cybersecurity capabilities. Mistral claims to have designed the Large 4 to resist indirect injection attacks and reports a resistance rate of 93.3% on Lakera’s B3 benchmark. The company also notes that the model maintains a high rejection rate of malicious requests despite its potential offensive capabilities, a combination that should allow its use in sensitive security environments.
Beyond performance, Mistral is committed to the open nature of Large 4. Once the models are available, companies and administrations will be able to adapt and deploy them on their own infrastructure, without depending exclusively on Mistral servers. For French society, this possibility constitutes an argument in favor of digital sovereignty, in particular for organizations that want to keep their data and artificial intelligence systems under their own control.
Mistral does not consider Large 4 to be a result. The company says its retraining is still ongoing and the model has significant room for improvement thanks to increased computing capabilities. The Big 4 should also serve as the basis for a new generation of specialized models developed for different sectors and uses.
