Mistral Saba

Mistral Saba is Mistral AI's first regional language model, a 24-billion-parameter LLM trained on curated datasets from the Middle East and South Asia rather than a generic Western-heavy mix. It targets Arabic interactions first and, through cultural cross-pollination in its training data, also handles Indian-origin languages such as Tamil and Malayalam.

Despite its modest size, Saba is designed to beat larger general-purpose models on regional content while running fast: over 150 tokens per second on a single GPU, with deployment options through Mistral's API or on-premise in a customer's own environment. It can also serve as a base for fine-tuned, domain-specific regional models.