Mistral Large 4 offers European open-weight alternative for coding and security

The release of Le Chonk

France’s Mistral recently released the open weights for its Mistral Large 4 model, which the company has nicknamed Le Chonk. Open weights mean the trained numerical parameters are public, allowing anyone to download, host, and fine-tune the model on their own hardware. This differs from standard open-source software because it provides users with practical control over their data and prevents external API shutoffs.

The model is designed with a mixture of experts architecture, which means only a portion of the network runs during each step of generating an answer. While it has 1 trillion total parameters, it uses about 49 billion active parameters per token. This design makes the inference process cheaper than the headline size might suggest while still maintaining high performance for complex tasks.

Mistral claims that three extra weeks of reinforcement training on difficult coding and security tasks significantly improved the model's capabilities. The model is natively multimodal and features a 1M token context window. It also offers three distinct reasoning effort settings, allowing users to choose between low, medium, and high intensity for different types of prompts.

Benchmarking the claims

Mistral reports that Le Chonk matches the best Chinese open models in agentic coding and cybersecurity. However, independent tests from Artificial Analysis and Vals.ai provide a more nuanced picture of these claims. While the model shows solid performance in several areas, some of the company's headline figures have not been fully reproduced by third-party testers.

For example, Mistral claimed a score on Terminal-Bench 4.0, but independent runs showed results between 25 to 26 percent. Similarly, the model scored 15 percent on FrontierCode 2.0 in independent tests compared to the company's claimed 19.7 percent. Despite these discrepancies, the model remains competitive, particularly on the Vibe Code Bench where it scored 81.56 percent.

On the composite Coding Agent Index, the model is currently tied with China's GLM-5.3. In terms of efficiency, Mistral claims the model requires about 47 percent fewer tokens per task than GLM-5.3 for comparable results. If these efficiency claims are verified, the model could offer significant cost savings for high-volume enterprise users who need to minimize their bills.

Sovereignty and security

A major focus of this release is the model's utility for cybersecurity defenders. It is pitched as a tool for finding bugs, triaging alerts, and writing patches within private networks. Because the weights are open, organizations can run the model entirely offline, which is critical for environments that cannot send sensitive code to a foreign API.

This capability is particularly relevant given recent reports of AI models escaping test sandboxes or being used for vulnerability hunting. Open weights are dual-use technology, meaning they can be used by defenders to secure systems or by attackers to find flaws. Experts suggest that monitoring and access controls remain essential regardless of where a model is trained or hosted.

The European angle is also significant for government procurement. Mistral trained the model on 3,800 Nvidia Grace Blackwell GPUs located in European data centers. This provides a sovereign option for entities that are hesitant to depend on US-based closed APIs or Chinese open-weight models, changing the conversation around AI infrastructure and data residency.

Hardware and licensing

While the weights are open, Mistral Large 4 is not intended to run on standard consumer hardware like a laptop. Because of its trillion-parameter scale, the official deployment recipes target 8 GPU servers. This means that while the model is accessible for private use, it requires significant infrastructure to operate effectively.

The model ships under the Mistral Ecosystem Open-Weight License. This allows for free use and fine-tuning, but it requires a separate commercial license for anyone offering the model as a service. This distinction ensures that while the weights are accessible for private use, the company can still manage the commercial distribution of the model as a product.

The competitive landscape

Until this release, Chinese labs like DeepSeek and Z.ai have largely set the pace for open-weight models in coding and reasoning. Mistral's goal with Le Chonk is to narrow that gap by offering a Western alternative that matches those models on specific workloads while maintaining high efficiency. It does not reset the entire field, but it establishes a new baseline for European AI capabilities.

The model is natively multimodal and offers three reasoning effort settings. It also features a 1M token context window, which allows it to process large amounts of information at once. These features combined with the open-weight availability make it a versatile tool for developers and security professionals.

Moving forward, the focus will be on whether independent tests can confirm Mistral's efficiency and cybersecurity claims. If the model holds up as a practical alternative for enterprise coding, it could become a staple for organizations prioritizing data sovereignty. For now, it serves as a significant milestone in the race to provide high-performing, hostable AI models outside of the major US cloud providers.