The models provided by Mistral AI, the latest tricolor unicorn since its December fundraising, were able to convince. Many companies are already launching experiments based on its technologies.
The startup is also increasing its partnerships, such as with Capgemini earlier in February. This month, Mistral is announcing in quick succession the availability of an additional model, Mistral Large, of its Chat assistant, as well as the signing of a partnership with Microsoft.
Mistral AI in MaaS on Azure
Already closely linked to OpenAI, the Redmond firm does not intend to limit its options. The partnership covers several areas. Mistral AI teams will thus have access to the Azure AI calculation infrastructure for training and inference of their models.
On a commercial level, their agreement also provides for the provision of French generative AI models in MaaS (Models as a Service) mode from Azure AI Studio and the Azure Machine Learning catalog.
“The possibility of using Microsoft Azure Consumption Commitment (MACC) to purchase Mistral AI models is available today,” the American firm further specifies in a press release dated February 26.
R&D partnership with Microsoft
Finally, the partnership covers AI R&D. In terms of research, Microsoft and Mistral AI will explore a collaboration around learning models specific to certain customers, including European public sector workloads.
But Microsoft also gets the scoop on Mistral AI’s latest LLM, Mistral Large, first available on Azure AI – and on its publisher’s platform. Called “Au Large” in French, the latest model of the hexagonal unicorn boasts performances that place it not far from GPT-4, OpenAI’s flagship.
Mistral Large also means mastering several languages (English, French, Spanish, German and Italian), “with a nuanced understanding of grammar and cultural context.” The developers add that thanks to its context window of 32K tokens, Large “allows you to find information precisely in large corpora of documents.”
Le Chat, Mistral’s VRP assistant
The designers also insist on the integration of an optional secure mode allowing the moderation of sensitive content and on its ability to call functions. This capability, combined with the “JSON output” mode of the Mistral platform, “allows the development of applications linked to existing systems, and the modernization of technology stacks on a large scale.”
But to make its technologies even more widely known, Mistral AI learned from its American competitor. OpenAI was able to make ChatGPT a formidable demonstrator and a formidable VRP. The startup is therefore launching Chat.
The conversational interface constitutes “an assistant that presents what can be built with our models. Chat and our models can be deployed anywhere, including in complex corporate environments,” underlines the unicorn.