French AI powerhouse Mistral AI has officially launched the public preview of its latest flagship model, Mistral Large 4 (ML4), codenamed ‘le Chonk’. Available starting today, users can access the model through an API on Mistral Studio. The company has announced that the model’s weights will be made publicly available for download by the end of this month.
‘Europe’s Strongest Open Model’
Mistral AI claims that ML4 is the most powerful open-weight model developed in Europe, rivaling and even surpassing top global open-source models in various benchmarks. The company asserts that ML4 achieves state-of-the-art performance across critical enterprise workloads, including cybersecurity, finance, and legal applications. In certain specialized domains, such as visual localization, ML4 is reported to outperform even current closed-source, cutting-edge models.
In cybersecurity, finance, and legal key enterprise workloads, it achieved SOTA among all open models.
In some specific domains (e.g. visual localization), it goes even further, outperforming current frontier closed models.
Advanced Capabilities and Training
ML4 is described as a native multimodal model boasting 1 trillion total parameters and 49 billion active parameters. Its training data is noted for its strong multilingual capabilities, covering over 160 natural languages, including all official languages of the European Union. This extensive language support ensures broad applicability and inclusivity.
The model was trained from scratch on Mistral AI’s proprietary infrastructure in Europe, utilizing 3,800 NVIDIA Grace Blackwell GPUs. Significantly, Mistral AI emphasizes that its European deployment is fully operated end-to-end and independently by Mistral AI, adhering strictly to EU laws and regulations, and remaining free from the influence of any third-party digital service providers.
Red-Teaming and Security Focus
Mistral AI is currently engaged in rigorous red-teaming efforts for ML4, collaborating with leading cybersecurity firms, trusted partners, and government regulators. These partners have been granted special access to the model for testing in real-world business scenarios. This version of the model features relaxed content safety policies and enhanced cybersecurity capabilities, reflecting a strong commitment to responsible AI development and security robustness.
Pricing and Availability
The public preview API for Mistral Large 4 is now accessible via Mistral Studio. The pricing is set at $1.36 per million input tokens and $4.18 per million output tokens. The full release of the model weights is anticipated by the end of the current month, further cementing Mistral AI’s position in the open-source AI landscape.
