On October 6, French enterprise Mistral AI officially launched the public preview of its Mistral Large 4 model. Developers often refer to this advanced artificial intelligence simply as ML4. Furthermore, the company gave it the intriguing codename “le Chonk.” Starting today, users can actively explore the preview API through Mistral Studio. Additionally, the developers plan to release the official model weights for download by the end of this month.
Currently, the organization is collaborating with top cybersecurity agencies. Trusted partners and government regulators are also participating in this vital effort. Together, they are conducting rigorous red-teaming tests on the model within real-world business environments. Consequently, these collaborating parties will receive exclusive access for specialized testing. This specific test version slightly relaxes standard content safety restrictions. Simultaneously, it significantly enhances capabilities for advanced cybersecurity offense and defense techniques. You can explore more details directly from the official Mistral Large 4 announcement.
Groundbreaking Model Architecture
ML4 operates as a native multimodal model featuring one trillion total parameters. It also utilizes an impressive 49 billion active parameters during processing. Mistral confidently asserts that its practical performance easily rivals the world’s leading open-source models. Moreover, it substantially outperforms all open-weight models developed across Europe and the United States. The system achieves state-of-the-art results in critical enterprise workloads. These demanding sectors include cybersecurity, finance, and complex legal analysis. In certain specialized domains like visual grounding, its capabilities extend even further. Surprisingly, it even surpasses several prominent closed-source frontier models.
European Infrastructure and Multilingual Support
The company trained ML4 entirely from scratch within its proprietary European data centers. They powered this massive undertaking using 3,800 Nvidia Grace Blackwell GPUs. Consequently, Mistral manages the entire European deployment strategy completely independently from end to end. This localized approach strictly complies with all European Union laws and regulations. Therefore, the system operates completely free from reliance on external digital service providers. The training data for ML4 possesses remarkable multilingual characteristics. Specifically, the corpus encompasses over 160 distinct natural languages. Ultimately, it fully supports every official language recognized within the European Union.
Competitive API Pricing
Regarding pricing, ML4 remains highly competitive for enterprise adoption. The company charges $1.36 per million input tokens. Meanwhile, the cost for output generation stands at $4.18 per million tokens. Ultimately, this aggressive pricing strategy positions the model as a highly attractive option for developers.
