The next border for models of large languages (LLMS), one of the key technologies underlying the boom of generating AI tools, could be geographic. Monday, AI startup based in Paris Mistral – Who is in the running to compete with Anthropic and Openai, based in the United States – publishes a slightly different model from its usual LLM.
Named Mistral Saba, the new custom -made model is designed to approach a specific geography: Arab speaking countries. Mistral Saba’s objective is to excel in Arab interactions.
Mistral Saba is a relatively small model with 24 billion parameters. As a reminder, fewer parameters lead to better performance with lower latency. But more parameters generally mean smarter responses, even if it is not a linear correlation.
Mistral Saba is of size comparable to Mistral Small 3Its small model for general use. But, according to the own tests of Mistral, Mistral Saba works much better than Mistral Small 3 when managing the Arab content.
As an interesting secondary effect, due to the cultural cross -pollination between the Middle East and South Asia, Saba also works well with languages of Indian origin, by Mistral – in particular Indian languages from the south, such as Tamil and Malayalam.
The new model represents an interesting strategic decision for the French AI giant, showing increased accent on the Middle East. Mistral said he expects the model to help him gain land among customers in the region.
As a standard model, Mistral Saba could be used for conversational support or generation of content in Arabic which seems more natural and relevant. It can also be used as a basis for certain refined models for internal use cases, said the company.
Last week, Mistral used the AI action summit to demonstrate that it becomes serious in business. Although the company has already collected large sums of international investors, many of its foreign donors are based in the United States – investors such as Lightspeed Venture Partners, Andreessen Horowitz and Salesforce Ventures.
Due to the changing geopolitical landscape, Mistral could potentially accommodate investors in the Middle East in its next financing round. It would be a way to collect more money to remain relevant in AI race at a technical level, while positioning itself as the international alternative to American and Chinese AI companies. The new Mistral model, Saba, could therefore contribute to this effort to collect potential funds.
Mistral Saba is accessible via the Mistral API. It can also be deployed on site, which could be a strong sale argument for companies working in sensitive industries, such as energy, finance or health care.
Due to the European roots of the company, since the release of the original Mistral 7B model of open weight, he often reiterated that he was taking the multi-language support seriously. The release of Saba is a continuation of this positioning. And Mistral said he will turn his attention to other regional languages on the road.
