Mistral AI launches Mistral Large and a partnership with Microsoft

Background

Mistral Large was the company's flagship proprietary model, and the release also came with a major cloud partnership. ^[1][2]

What happened

On 26 February 2024 Mistral AI released Mistral Large, its flagship proprietary model, along with Mistral Small. ^[1] Mistral claimed that Mistral Large was the second-ranked model available through an API after GPT-4. It has a 32K context window. ^[1] The company also launched its Le Chat assistant in beta on the same day. ^[1]

The same day Microsoft announced a multi-year partnership with Mistral AI. Mistral Large was made available first on Azure AI and on Mistral's own platform. ^[2]

The context window is the amount of text a model can take into account in one exchange, measured in tokens. A flagship model is the largest and most capable model in a company's range, usually the one it promotes most. Azure AI Studio is a Microsoft Azure service where developers can build with and deploy models from a catalogue, and Azure supercomputing refers to the large clusters of accelerator chips that cloud providers use to train and run big models.

Key details

The partnership covers Azure supercomputing for training and inference, which are the two stages of building and using an AI model. It also covers distribution through Azure AI Studio and joint research and development, including European public-sector workloads. ^[2]

Wikipedia reports that Microsoft invested $16 million in the company. ^[3]

A joint research and development agreement means that two companies work together on new technology, and such agreements often include shared engineering work rather than only the sale of a product. Partnerships of this kind can give a young company access to infrastructure it could not build alone.

Impact

The partnership gave the company a major cloud partner for distributing its models. ^[2][3] The flagship model offered through a cloud provider, together with the Le Chat assistant launched in beta, gave the company two ways to reach users. ^[1][2]