Mistral AI announces the availability, on Hugging Face, of the weights for Mistral Large 4, nicknamed ‘Le Chonk’. This multimodal model, capable of processing multiple types of content, has 1 trillion parameters, of which 52 billion are active. The company claims leading results on several benchmarks, standardized performance tests, in programming.
The model is also accessible via Mistral’s API and can be tried in Vibe Code. Presented on October 6, Large 4 benefited from three additional weeks of reinforcement training, during which Mistral says it incorporated feedback from customers and partners. According to the company, this phase improved its performance and added three levels of reasoning: low, medium, and high.
Mistral presents Large 4 as the best open model developed in the West. It also ranks first globally among open models for agentic software development, according to Artificial Analysis’s Coding Agent Index. Agents are programs capable of chaining actions autonomously.
In agentic programming, Large 4 rivals GLM-5.3 while using, on average, 47% fewer tokens per task, according to Mistral. Tokens are the units of text processed by the model, and their number directly influences the cost of use. On Terminal-Bench 4.0, a benchmark evaluating agents capable of performing programming operations in a terminal, its score reportedly increased from 28% to 41%, thus reaching the level of GLM-5.3.
Cybersecurity among the targeted applications
Mistral also highlights the model’s capabilities in cybersecurity. It is notably intended to assist teams in code analysis, threat detection, and examination of IT infrastructures. The release of the weights offers companies and administrations the possibility of running the model in their own environment, without depending exclusively on a remote service. This option may interest organizations that must retain control of their data and systems, particularly in regulated sectors, with Mistral presenting it as a way to strengthen control over AI usage.
The model scores 51% on the Artificial Analysis Cyber Index, which notably evaluates its ability to identify, reproduce, and correct computer vulnerabilities. Mistral also announces an average refusal rate of 96.6% for malicious cyber requests, according to several specialized test sets.
Large 4 was trained on 3,800 NVIDIA Grace Blackwell GPUs, graphics processors, in Mistral’s European data centers. The company highlights the optimization of its computing resources to explain performance achieved with financial means lower than those mobilized by some competitors. Arthur Mensch, co-founder and CEO of Mistral, recalls that the company has raised about $6 billion in three years. The weights are published under the Mistral Ecosystem Open-Weight License, a license derived from the MIT license that governs the redistribution and commercial exploitation of the model.
Outlets in industry and finance
The envisioned applications go beyond software development. In industry, Mistral cites transforming 2D plans into 3D models usable in computer-aided design (CAD) software. In finance, Large 4 could process complex spreadsheets, analyze large volumes of data, and exploit financial publications.
Its multimodal capabilities could also be used to analyze satellite and aerial images, for example to detect fires, monitor crop status, or identify risks to electrical infrastructure.
