Devstral 2 (123B)

Mistral AI conducts post-market monitoring to track the performance, safety, and compliance of its AI models, including open-source releases, after deployment. This involves continuous evaluation of model outputs, user feedback, and incident reporting to detect and address risks, biases, or unintended behaviors. Mistral actively monitors for compliance with ethical guidelines and regulatory frameworks, such as the EU AI Act, ensuring responsible AI use.

Mistral models have a finite context window (e.g., 32k tokens for some versions). This means they may struggle with very long documents or conversations, potentially losing track of earlier details. While strong at many tasks, the models can make logical errors or oversimplify nuanced reasoning, especially in highly technical or abstract domains. Mistral models are trained on data up to a specific cutoff date (e.g., November 2024 for some versions). They may not have real-time or post-cutoff knowledge unless fine-tuned or augmented with external tools. While multilingual, performance is generally stronger in high-resource languages (e.g., English, French) compared to low-resource languages.