Mistral AI released the weights of Large 4, an artificial intelligence (AI) model with one trillion parameters that can be downloaded and run by companies and governments on their own systems. Nicknamed “Le Chonk”, the AI was designed for tasks such as identifying vulnerabilities in software, analyzing digital threats, and helping respond to cyberattacks.
The launch took place this Thursday (29) and reinforces the French company’s bet on open-weight models, which allow greater control over the technology. The files are available on Hugging Face, and the tool can also be accessed through the Mistral API.
The goal is to reduce dependence on AI services controlled by external providers. This is especially relevant to organizations that handle sensitive information and need to analyze security incidents without transferring their data to third-party platforms.
How does Mistral Large 4 work?
- Large 4 has one trillion parameters in total, of which 52 billion are activated for each task, according to Mistral. The model is multimodal, meaning it can process information in different formats, including text and images, and generate responses and code;
- The company says the AI was trained on its own infrastructure in Europe, using 3,800 NVIDIA Grace Blackwell GPUs. The system was also developed for programming, executing tasks with AI agents, and understanding multimodal information;
- The model had already been introduced in testing version in early October. Since then, Mistral expanded the training and incorporated evaluations and feedback from customers and partners to enhance the performance of the version made available to the public.
AI Promises to Help Find Security Flaws
Cybersecurity is one of the main focuses of the launch. Large 4 was developed to help technical teams locate vulnerabilities, reproduce flaws to verify whether they are real, and produce fixes for the identified problems.
According to Mistral, the model scored 82% on the Artificial Analysis Cyber Index test, which evaluates the ability to reproduce a real vulnerability in open-source software and then fix it. The company also says that the AI solved 93% of the Cybench challenges, a set of 40 exercises inspired by digital security competitions.
The results are published by the developer itself and should be interpreted within the specific conditions of each evaluation.
In addition to identifying flaws, Mistral points to applications in malware analysis, vulnerability prioritization, and the creation of rules to detect threats.
The company also highlights its performance in programming. In Artificial Analysis’s Coding Agent Index, Large 4 achieved a combined score of 49.8%, according to data published by the developer. Mistral says the result places it ahead of other open-weight models in programming tasks performed by AI agents.
TAG42@@
Large 4 was developed to help technical teams locate vulnerabilities, reproduce flaws to verify whether they are real, and develop fixes for identified issues – Image: T. Schneider/Shutterstock
Read more:
- This is Europe’s answer to ChatGPT
- “Vassal state”: CEO of an OpenAI rival issues a stern warning to Europe about AI
- Mistral partners with Airbus and BMW to expand AI in manufacturing
Why release the model’s weights?
Weights are the numerical values learned during training that enable the model to generate its responses. By releasing these files, organizations can run Large 4 on their own infrastructure instead of relying exclusively on an external service.
This allows greater autonomy over data, tool configuration, and security policies in place. In a critical situation, for example, a company can use the model to investigate an attack without depending on the availability or restrictions of a third-party API.
In an interview with TechCrunch, Pierre Stock, Mistral’s vice president of science, said that making the weights available enables companies and governments to access advanced AI resources to defend themselves against attacks. According to him, the model can help identify, coordinate, and respond to major security incidents in real time.He also advocated broadening access to these capabilities to strengthen the digital protection ecosystem.
The strategy puts Mistral in competition with companies such as OpenAI and Anthropic, whose most advanced models are offered primarily through services controlled by the developers themselves.
Opening up also brings risks
The ability to run a model without relying on an external provider comes with a trade-off: once the weights are made available, it becomes more difficult to control who uses the technology and for what purposes.
The same ability to analyze vulnerabilities that helps security teams may also appeal to malicious agents seeking to exploit flaws in systems. In addition, independent operation allows organizations to adapt the model to their own needs, but it also means that restrictions imposed in a hosted service do not necessarily carry over to all modified versions.
Stock acknowledged the existence of risks associated with the advancement of AI and autonomous agents, but said there is no need to amplify fear around the technology. For him, security is a challenge that can be faced.
Mistral maintains that Large 4 was developed to strengthen cybersecurity, with the ability to perform technical tasks that more restrictive systems may refuse, even when used in legitimate security research.
With the launch, the company expands its offering of advanced models that can be run and adapted by users themselves.The challenge will be to balance access to powerful digital defense tools with the risks of their improper use.
Rodrigo Mozelli is a journalist graduated from São Paulo Methodist University (UMESP) and is currently a writer at Olhar Digital.
