Mistral AI has made its latest AI model, Mistral Large 4, nicknamed ‘Le Chonk’, available for use on organisations’ own infrastructure. The multimodal model has 1 trillion parameters and, according to the French company, was trained entirely in European data centres. Mistral claims the model ranks among the best-performing open AI models for software development and defensive cybersecurity. Organisations can download the model, run it independently and adapt it to their own applications.
The model was released as a public preview earlier this month. Since then, Mistral AI has trained it further using reinforcement learning. According to the company, the model achieves the highest scores on software development benchmarks, including FrontierCode 2.0 and Terminal-Bench 4.0. On the Artificial Analysis Coding Agent Index, Mistral AI says Mistral Large 4 achieves scores comparable to those of GLM-5.3, while using an average of 47 per cent fewer tokens per task.
Three reasoning modes
Developers can choose from three reasoning modes to control the balance between performance and compute costs: Low for simple tasks, Medium for everyday use and High for complex problems. Mistral AI says this gives developers more control over token consumption per task.
Defensive cybersecurity
Mistral AI reports that the model has been specifically trained for defensive cybersecurity. According to the company, it scores 51 per cent on the Artificial Analysis Cyber Index for detecting, reproducing and fixing vulnerabilities in source code. To limit misuse, the model includes guardrails. On benchmarks such as JailbreakBench and StrongREJECT, Mistral AI says the model refuses an average of 96.6 per cent of malicious or offensive cyber requests. The company describes this as the highest score among open-weight models.
According to Mistral AI, organisations testing or integrating the model include Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, Fortinet, Orange Cyberdefense, Rubrik, Zscaler and the European Union Agency for Cybersecurity, ENISA. The company is also introducing a separate variant for penetration testing and red-teaming for selected enterprise partners. This variant is only available under strict conditions.
Self-hosting and customisation
Mistral AI states that the model was developed entirely in-house and trained on European infrastructure. Organisations can run it locally, customise it via Mistral Forge and manage it without depending on an external API. The company is therefore positioning the release as a foundation for digital sovereignty. According to Mistral AI, organisations including the Dutch bank ABN AMRO, La Banque Postale, Belfius, Crédit Agricole, ASML and the European Space Agency (ESA) have confirmed their involvement in the initiative or their use of the model.
Availability and licensing
Mistral Large 4 is available immediately via Hugging Face, Mistral Studio, the Mistral AI API and platforms such as OpenRouter, Microsoft Foundry and NVIDIA NIM. The model is covered by the Mistral Ecosystem Open-Weight licence, which permits free use, modification and commercial deployment. Parties offering the model as a commercial Model-as-a-Service offering require a separate licence. IT service providers that want to host the model on behalf of clients should therefore check whether their service falls under this exception.
This article was originally published in Dutch on MSP Business. Read the original →
