Large language models (LLM) are one of the key technologies supporting the booming development of generative artificial intelligence tools, and its next frontier may be geographical location. On Monday, Paris-based artificial intelligence startup Mistral released a model that’s a little different from its usual LLM.
The new custom training model, named MistralSaba, is designed to address a specific geography: Arabic-speaking countries. MistralSaba's goal is to excel in Arabic interactions.
MistralSaba is a relatively small model with 24 billion parameters. It should be reminded that the fewer parameters, the better the performance and the lower the latency. But more parameters usually mean smarter answers, although this is not a linear correlation.
MistralSaba is comparable in size to MistralSmall3 (its universal small model). However, according to Mistral's own testing, MistralSaba performs much better than MistralSmall3 when it comes to handling Arabic content.
An interesting utility is that due to the cultural exchange between the Middle East and South Asia, Mistral believes that Saba also handles Indian languages well, especially South Indian languages such as Tamil and Malayalam.
The launch of this new model is an interesting strategic move for the French AI giant, signaling a greater focus on the Middle East. Mistral said it hopes the model will help it gain traction among customers in the region.
As a ready-made model, MistralSaba can be used for Arabic conversation support or content generation that sounds more natural and relevant. The company says it can also serve as the basis for some fine-tuned models for internal use cases.
TechCrunchDisrupt2025
Disrupt2025 will be here before you know it! Get your tickets now at the lowest prices all year. From AI and startups to space, fintech and IPOs – experience game-changing insights through five main stages, breakout sessions, roundtables, unparalleled networking and more.
San Francisco, CA | October 27-29
Register now
Last week, Mistral used the AI Action Summit to show it was serious about its business. While Mistral has raised significant funding from international investors, many of its foreign backers are based in the United States, such as Lightspeed Venture Partners, Andreessen Horowitz and Salesforce Ventures.
Due to the changing geopolitical landscape, the company is likely to welcome Middle Eastern investors in an upcoming round of funding. This would be a way to raise more funds to continue participating in the AI race on a technical level and position itself as an international alternative to U.S. and Chinese AI companies.
Mistral's latest model, the Saba, could contribute to this potential funding effort.
MistralSaba is accessible through Mistral's API. It can also be deployed on-premises, which can be a strong selling point for companies in sensitive industries such as energy, finance or healthcare.
Due to its European roots, the company has often reiterated its strong emphasis on multi-language support since the release of the original open Mistral 7B model. The release of Saba is a continuation of this positioning. The company said it will turn its attention to other regional languages in the future.