Mistral AI releases weights for its new Large 4 model

The launch of Mistral Large 4

Mistral AI has officially released the weights for its newest model, Mistral Large 4, also known as "Le Chonk." By releasing the weights, the company allows developers and organizations to download and run the model on their own hardware rather than relying solely on a remote service. This move is intended to give users more control over their data and how the AI is deployed within their specific environments.

The model is multimodal, meaning it can process different types of information, and it was made available on Hugging Face. It is also accessible through the Mistral API and can be tested within the Vibe Code platform. Mistral positions this release as the best open-source model developed in the Western world.

The company spent three additional weeks on reinforcement learning before the launch. During this period, the team integrated feedback from various partners and clients to refine the model's behavior. This extra training resulted in three distinct levels of reasoning: low, medium, and high.

Efficiency in software development

One of the primary focuses for Large 4 is its performance in agentic software development. According to the Coding Agent Index by Artificial Analysis, the model ranks at the top of the world for open-source models in this category. It is specifically designed to help with complex programming tasks that require an agent to perform multiple steps.

Mistral highlights that Large 4 can rival the GLM-5.3 model while being significantly more efficient. The company reports that the model uses an average of 47% fewer tokens per task than GLM-5.3. This is a major economic advantage for businesses because token consumption directly determines the cost of running AI agents at scale.

Testing on Terminal-Bench 4.0, which evaluates an agent's ability to perform programming operations in a terminal, showed significant progress. The model's score rose from 28% to 41% on this benchmark. This jump allows it to reach the same performance level as GLM-5.3 while maintaining a lower cost profile for the user.

Security and infrastructure

Beyond coding, Mistral is positioning Large 4 as a tool for cybersecurity teams. The model is designed to assist with code analysis, threat detection, and the examination of IT infrastructures. It aims to provide a layer of automated oversight for identifying vulnerabilities within a company's digital systems.

The model's capabilities were tested on the Artificial Analysis Cyber Index, where it scored 51%. This index measures how well a model can identify, replicate, and correct various computer vulnerabilities. These results suggest that the model can be a practical assistant for security professionals looking to harden their systems.

Mistral also emphasized the model's safety profile regarding malicious requests. In several specialized test sets, the model showed an average refusal rate of 96.6% when faced with cyber-related malicious prompts. This high refusal rate is intended to ensure that the model is not easily manipulated into performing harmful actions.

Training and licensing details

The development of Large 4 required significant computing power, utilizing NVIDIA Grace Blackwell GPUs. These were located in Mistral's data centers across Europe. The company noted that it optimized its computing resources to achieve these results with lower financial investment than some of its competitors.

To put the company's scale into perspective, Arthur Mensch, the co-founder and CEO of Mistral, noted that the firm raised approximately 6 Md$ over the last three years. This funding has supported the development of their high-performance models and the infrastructure needed to train them.

The model is distributed under the Mistral Ecosystem Open-Weight License. This specific license is a derivative of the MIT license and includes specific rules regarding the redistribution and commercial exploitation of the model. This framework allows for commercial use while maintaining certain conditions for the Mistral ecosystem.

Industrial and financial uses

Mistral envisions applications for Large 4 that extend far beyond simple software engineering. In the industrial sector, the model could be used to transform 2D plans into 3D models for use in computer-aided design (CAD) software. This could streamline the manufacturing and construction processes for various engineering firms.

The finance industry is another target for the model's capabilities. It is expected to handle complex spreadsheets, analyze massive volumes of data, and extract insights from financial publications. These tasks require the high-level reasoning and data processing power that the 1 trillion parameters provide.

Finally, the model's multimodal nature allows for the analysis of satellite and aerial imagery. Potential use cases include detecting fires, monitoring the status of crops, or identifying risks to electrical infrastructure. These applications demonstrate the model's versatility across different sectors of the economy.