Mistral AI unveils Mistral Saba for Arab world & South Asia
New Mistral model has enhanced Arabic and South Asian language capabilities
#France #LLMs - Mistral AI, the pioneering French AI startup, has announced Mistral Saba, the developer’s first regional language model. Designed to enhance AI fluency in Arabic and South Asian languages, Mistral Saba is a 24 billion parameter model trained on curated datasets from the Middle East and South Asia. Developers claim the new model offers faster and cost-effective deployment, and has showed to be more efficient than models over five times its size in benchmarking. The release of Mistral Saba could signal new interest from global AI model developers to create AI models that can compete effectively in the Middle East’s growing GenAI market.
SO WHAT? - It is rarely that a global large language model, typically strongest in Western languages, is trained to be more Arabic-centric. The new Mistral Saba is one of an increasing number of custom-trained models from the developer to serve specific geographies, markets, and customers. Mistral Saba now provides another option for GenAI app developers in the Middle East looking for the right Arabic LLM.
Some key points regarding the announcement of Mistral Saba:
Mistral AI has launched Mistral Saba, a 24-billion parameter AI model trained specifically for Arabic and South Asian languages. It is the developer’s first regional language model.
According to Mistral AI, the model supports Arabic, and has also been trained on many Indian-origin languages. It is particularly strong in the South Indian-origin languages of Tamil and Malayalam. No information is available on the origin of data, the size of curated datasets or tokenisation process.
According to Mistral AI’s evaluation, the model provides more accurate responses than much larger models while being faster and more cost-effective.
Mistral Saba is available via API and on-premise deployment, ensuring security and control for businesses.
Designed for applications such as conversational AI, industry-specific expertise, and culturally nuanced content creation. Again, what data informed the model’s cultural bias has not been made public.
Mistral Saba can be fine-tuned for domains like energy, finance, and healthcare, enhancing sector-specific AI capabilities, and according to Mistral AI, it can do this all within the context of Arabic language and culture.
Mistral Saba 24B delivers speeds exceed 150 tokens per second on single-GPU systems, making it lightweight and efficient.
ZOOM OUT - Mistral AI set records in early 2023 by raising $118 million just a month after its founding, marking the largest seed round in European history. Valued at $2 billion post-funding, Mistral AI launched its first large language model, Mistral 7B, a few months later. The new model outperformed larger competitors like Meta’s Llama 2 13B and led a new category of small language model, built for efficiency, ease of deployment and resource savings.
BENCHMARKS
LINKS
Blogpost with API and contact info (Mistral AI)