Mistral launches Large 4, its enterprise-grade code model
Mistral AI releases the weights of Large 4 on Hugging Face and strengthens its open model focused on software development and cybersecurity. The company highlights its performance, token efficiency, and applications in industry and finance.
Mistral AI has made available the weights of Mistral Large 4, its open multimodal model nicknamed "Le Chonk," on Hugging Face, with access also via the company’s API and in Vibe Code. Introduced on October 6, this model has 1 trillion parameters, including 52 billion active ones, and primarily targets software development, cybersecurity, and certain broader business uses. For companies, the challenge is simple: more control over execution, data, and costs, in a market where every token consumed ends up on the bill.
An open model designed to code, reason, and cost less
Mistral describes Large 4 as its best open model developed in the West. The company states it tops the Coding Agent Index by Artificial Analysis for agentic software development and competes with GLM-5.3 while using on average 47% fewer tokens per task. On Terminal-Bench 4.0, its score reportedly rose from 28% to 41% after three additional weeks of reinforcement training. In short, the model aims not only to be better but also less resource-hungry.
This additional training phase incorporated feedback from clients and partners, according to Mistral. The company says it added three levels of reasoning — low, medium, and high — to better tailor the model to use cases. This approach matters for technical teams, as the cost of an AI agent depends as much on response quality as on token consumption, those units of text generated or read by the model.
Cybersecurity becomes a very concrete selling point
Mistral also highlights Large 4’s cybersecurity capabilities. The model is intended to help analyze code, detect threats, and examine IT infrastructures. On the Artificial Analysis Cyber Index, it scores 51%, while Mistral reports an average rejection rate of 96.6% against malicious cyber requests, based on several specialized test sets. These figures should be interpreted within their benchmark scope but show where the company wants to position its product: in environments where mistakes are costly.
Publishing the weights also changes the deployment mode. Companies and administrations can run the model in their own environment without relying solely on a remote service. This choice is particularly appealing to regulated sectors that want to keep control over their data and systems. The weights are published under the Mistral Ecosystem Open-Weight License, a license derived from MIT that governs redistribution and commercial use.
A technological showcase for industry, finance, and infrastructure
The use cases mentioned by Mistral go beyond coding alone. In industry, Large 4 can transform 2D plans into 3D models usable in computer-aided design software. In finance, it is presented as capable of handling complex spreadsheets, analyzing large volumes of data, and exploiting financial publications. Its multimodality also opens the way to satellite and aerial image analysis to monitor crop conditions, detect fires, or identify risks on electrical infrastructures.
The model was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data centers. Arthur Mensch, co-founder and CEO, recalls that the company has raised about $6 billion in three years. Behind the technical discourse, the message is quite clear: Mistral wants to prove that a European player can keep pace with cutting-edge open models while maintaining a credible proposition for companies that first look at costs, then performance charts.
Key points
- Large 4 has 1 trillion parameters, including 52 billion active ones.
- Introduced on October 6, then published on Hugging Face.
- 47% fewer tokens, according to Mistral, compared to GLM-5.3.
- 51% on the Artificial Analysis Cyber Index.
- 3,800 NVIDIA Grace Blackwell GPUs used for training.
- Arthur Mensch mentions about $6 billion raised in three years.
By the numbers
- 1 trillion — parameters of Mistral Large 4, according to Mistral.
- 52 billion — active parameters of the model, according to Mistral.
- 47% — tokens saved on average per task, according to Mistral.
- 96.6% — average rejection rate of malicious cyber requests, according to Mistral.
- 3,800 GPUs — resources used for training in Mistral’s European data centers.