French AI startup Mistral has announced its foray into providing large language models (LLMs) that understand regional languages and their parlance, driven by rising demand from its enterprise customers. The company has released its first custom-trained regional language-focused model, Saba, which is a 24-billion parameter model trained on meticulously curated datasets from across the Middle East and South Asia.

Saba is designed to support use cases in Arabic and many Indian-origin languages, particularly South Indian-origin languages such as Tamil. This move marks a significant shift in Mistral's strategy, as the company acknowledges the importance of addressing every culture and language to make AI ubiquitous. According to Mistral, larger LLMs often fail to understand the usage of words in a certain language or lack understanding of the cultural background, leading to failure in servicing use cases in local languages.

Saba's custom training enables the model to grasp the unique intricacies and insights for delivering precision and authenticity. The model's support for multiple languages could increase its adoption, and its lightweight design makes it adaptable for a variety of use cases. Saba can be deployed via an API or locally on-premises, making it an attractive option for regulated industries such as finance, banking, and healthcare.

In benchmark tests, Saba outperforms several prominent LLMs, including Mistral Small 3, Qwen 2.5 32B, Llama 3.1 70B, and G42's Jais 70B. This demonstrates Saba's capabilities in understanding regional languages and its potential to drive adoption in underserved markets.

Analysts believe that Mistral's focus on regional language LLMs could help the company expand its revenue, driven by demand for localized AI solutions in sectors like finance, healthcare, and government. According to Charlie Dai, principal analyst at Forrester, "There's a growing market for regional LLMs like Saba, especially for enterprises needing culturally and linguistically tailored solutions. The market could be significant, driven by demand for localized AI in sectors like finance, healthcare, and government, potentially reaching billions as businesses seek to enhance customer engagement and operational efficiency."

However, Mistral is not the only model provider trying to capitalize on the regional language model trend. BAAI from China open-sourced its Arabic Language Model (ALM) in 2022, and DAMO of Alibaba Cloud open-sourced its PolyLM in 2023, covering eleven languages including Arabic, Spanish, and German. In the Middle East, regional public sector organizations have been attempting to create Arabic LLMs, such as the Saudi Data and AI Authority (SDAIA) that launched its LLM named ALLaM on IBM Cloud last year.

In South Asia, specifically in India, several startups have used Llama 2 to create regional language models, such as OpenHathi-Hi-v0.1 for Hindi, Tamil Llama, Telegu Llama, odia_llama2_7B_v1, and VinaLLaMA for Vietnamese. Despite the competition, Dai believes that high-quality, localized solutions will gain loyalty and market share in underserved areas, and regional business operations around the models are key to success.

Mistral's move into regional language LLMs marks a significant shift in the AI landscape, as companies increasingly recognize the importance of catering to diverse linguistic and cultural needs. As the demand for localized AI solutions continues to grow, Mistral's Saba model is well-positioned to drive adoption and revenue growth in underserved markets.