Microsoft allies with OpenAI rival Mistral AI

The partnership

Microsoft and the AI startup -- which has raised about $531 million and achieved a valuation of more than $2 billion since its founding last year -- will also collaborate on training purpose-specific models for certain customers, such as those in the European public sector, the cloud vendor said.

The investment in the upstart generative AI vendor, based in Paris, comes as Microsoft faces scrutiny from officials in the U.S., U.K. and European Union over its $13 billion investment in OpenAI. It was seen by some as an effort to address concerns about market dominance by allying with Europe's version of OpenAI as well.

For Mistral, the move comes on the same day the startup introduced a new AI chatbot, Le Chat, and language model, Mistral Large -- now available on the Azure and Mistral platform.

Mistral Large can be used for complex multilingual reasoning, according to the 10-month-old startup. It is fluent in English, French, Spanish, German and Italian. It also has a 32K-token context window -- the space a generative AI model uses to contain text or image data.

The context window of OpenAI's GPT-4 LLM is also 32K. Meanwhile, GPT-4 Turbo boasts a 128K context window.

Le Chat can use Mistral Large, previous model Mistral Small or prototype model Mistral Next. It also warns users when they're pushing the conversation in an inappropriate or unsafe direction.

Microsoft and open source

While Microsoft's inclusion of Mistral models on Azure was not unexpected, its investment in the French startup is surprising, Gartner analyst Arun Chandrasekaran said.

The tech giant said at its 2023 Microsoft Ignite conference in November that Mistral 7B, the startup's first foundation model, would be integrated into the Azure AI model catalog.

Microsoft has been clearly monitoring some of the emerging trends in the space and trying to craft their strategy to accommodate these emerging trends.

"Microsoft has been clearly monitoring some of the emerging trends in the space and trying to craft their strategy to accommodate these emerging trends," Chandrasekaran said.

One of those trends is the rise of small and midsize language models. Another trend is the growing popularity of open source models, such as Mistral's.

"Mistral in a relatively short time has emerged as one of the significant open source model companies in the space," Chandrasekaran said.

For example, the AI startup's Mistral 8x7B -- built for text generation, summarization, question and answering, and code generation -- is seen as comparable in capabilities and power to OpenAI's GPT-3.5 LLM.

"Not having these ... models natively on its platform was clearly a handicap for Microsoft, which they're now trying to address with the partnership and investment in Mistral," Chandrasekaran added.

And the new partnership reflects Microsoft's desire to ensure rivals don't overtake the momentum it has built in the generative AI market.

Microsoft's closest competitor in the generative AI race, Google, introduced its open source model Gemma on Feb. 21. Moreover, on Feb. 23, AWS, another Microsoft rival, revealed that two Mistral models -- Mistral 7B and Mistral 8x7B -- will soon be available on Amazon Bedrock.

Providing choices and regional opportunities

Microsoft is aware that the generative AI market is not monolithic in terms of the types of models enterprises are looking for, AI analyst Mark Beccue said.

"They know the market is going to want choices," he said, referring to Microsoft.

The partnership with Mistral also indicates that Microsoft is not solely wedded to OpenAI.

"They understand that not one size or one flavor fits all," Beccue said. "They're pursuing that for their customer base to make sure that they have those things to work with."

Moreover, the partnership helps Microsoft target public sector or government-affiliated agencies in Western Europe with the Mistral product, Chandrasekaran said.

With stricter AI regulations coming in Europe, Mistral has shown itself to be an AI company that takes a responsible approach, with safety measures already built in to guard against LLM hallucinations and bias, Beccue said.

"I don't see Mistral needing guardrails that OpenAI models do," he said. "They're bright. A lot of people are picking them very quickly on the enterprise side."

For Microsoft, the challenge will be how to monetize open source models, Chandrasekaran said. 

Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.