Paris-based AI startup Mistral has released Mistral Saba, an LLM the startup describes as "the first of its specialized regional language models." Mistral Saba, a 24 billion-parameter model, was tailored to excel in providing relevant responses to users in the Middle East and South Asia and support regional languages such as Arabic and many Indian-origin languages, especially those from South India like Tamil or Malayalam.

Mistral Saba is about the same size as the recent Mistral Small 3, but the differences in performance on Arabic-specific benchmarks between both models are noticeable. In particular, Mistral Saba outperforms Mistral Small 3 across the board, along with other popular LLMs such as Llama 3.3 70B, Cohere's Command R, and GPT 4o mini. Notably, Mistral Saba underperforms in English language benchmarks, as can be seen by the fact it was surpassed by Mistral Small 3 in the English MT-Bench Dev and MMLU evaluations.

According to Mistral AI, Mistral Saba is available as an API or can be deployed on-premises on settings as small as single-GPU systems to ensure an additional layer of security and privacy, which is especially well-suited for applications in sensitive industries like energy, finance, and healthcare. Early use cases include conversational assistants, domain-specific expertise, and culturally relevant content creation.

In addition to seeing Mistral Saba as the product of the startup's ongoing commitment to providing multi-language support, some speculate that this move could be part of Mistral AI's strategy to establish itself as a truly international alternative to China or US-based companies that have a narrower focus which prioritizes their corresponding domestic markets.