In February, Microsoft announced a new partnership with French startup Mistral AI. The multi-year agreement allows customers of Microsoft’s Azure AI services to access large language models created by Mistral AI on its Azure cloud servers.

The partnership began with Microsoft adding Mistral Large LLM to Azure AI customers. This week, Microsoft revealed that those customers can now also access Mistral Small LLM.

in a blog postMicrosoft reveals the features of Mistral Small and the types of AI tasks it can perform:

• A small model optimized for low latency: very efficient for high-volume workloads and low-latency workloads. Mistral Small is Mistral’s smallest proprietary model, it outperforms Mixtral 8x7B and has lower latency.
• RAG specialist: essential information is not lost in the middle of long contact windows. Supports up to 32K tokens.
• Strong in coding: code generation, review and comments with support for all mainstream coding languages.
• Multilingual by design: the best classroom performance in French, German, Spanish and Italian – in addition to English. Dozens of other languages ​​are supported.
• Effective safety rails baked into the model, with an additional layer of safety with a safe ordering option

Microsoft says access to the new LLM is available With an Azure subscription. Its customers can test Mistral Small using its Azure AI Studio services.

Just a reminder: Mistral AI was founded in 2023 by several former employees of Meta’s DeepMind division and Google. It raised $415 million last year. According to reports, Microsoft’s partnership with the startup also included a small investment in the company, but the actual financial numbers have not yet been released.

Microsoft has been a major investor and partner in a similar startup, OpenAI, for several years. Earlier this week, newly disclosed internal Microsoft emails showed that in 2019, the company fears that its main rival, Google, will beat them in the development of artificial intelligence. This concern led to Microsoft’s first investment in OpenAI later that year.