Last year, it announced in-depth cooperation with Qualcomm's computing platformMistral AI, earlier announced the creation of a high-performance artificial intelligence model for the Middle East and Southeast AsiaMistral SabaIt claims to be able to understand Arabic and various Indian languages, and it itself has a parameter scale of 240 billion sets.
Like other large-scale natural language models, Mistral Saba can not only be accessed through an API, but can also be installed on terminal devices for offline use, thereby ensuring data privacy during application.
Compared with other large-scale natural language models that cover a wide range of languages, Mistral Saba boasts a deep understanding of Middle Eastern and Southeast Asian languages, and can distinguish the cultural differences, professional terminology, and subtle differences in grammatical structure behind different languages.
As for the training data, it is derived from local data in the Middle East and Southeast Asia, allowing Mistral Saba to better meet the expected performance of localized language understanding and related applications.
In relevant test results, Mistral Saba's performance in large-scale, multi-task Arabic understanding even exceeded that of some models with a parameter scale of 700 billion groups. At the same time, it has lower inference latency and can even generate 150 tokens per second with a single set of GPU-accelerated computing, which is more conducive to deployment and application in environments with limited resources or usage efficiency in offline state.
Mistral Saba can also be fine-tuned to suit different needs and can be used in finance, energy, healthcare and other fields.
Mistral Saba performance comparison:
