Late on Monday night, Beijing time, MistralAI, which has become Europe’s leader in artificial intelligence less than a year after starting its business, sent out several messages in succession——Released the latest large model MistralLarge, launched the firstchatbot product LeChat, andReached a deeper cooperation agreement with Microsoft.

MistralAI was officially established in May 2023. At this time last year, the three young French people who founded this company were still working at Google DeepMind and Facebook.

By December last year, as Nvidia, Salesforce and a number of Wall Street capital waved checks and rushed to participate in the company’s financing worth more than 450 million euros (nearly 490 million US dollars), MistralAI’sThe market value has exceeded US$2 billion.


(The three founders of MistralAI, Guillaume Lample, Arthur Mensch, and Timothée Lacroix, are all only in their 30s)

Released the latest large model & the first chatbot

MistralAI announced on Monday that it has officially launched its latest flagship model, MistralLarge. The company claims that the model has "top-level reasoning capabilities" and can be used to handle complex multi-language reasoning tasks, including text understanding, transformation and code generation.

In the comparison of MMLU, the AI ​​understanding ability test benchmark,The score of MistralLarge is second only to GPT-4 and slightly better than Claude2 developed by Anthropic. As for Google's GeminiPro and Meta's LLaMA270B models, they were left behind.


(Source: MistralAI)

It is reported that this model has a context window of 32,000 tokens. According to the common algorithm that one token is approximately equal to 0.75 English words, MistralLarge can accurately extract information from documents containing nearly 24,000 English words at one time.

According to the company, the model isNative-level fluency in English, French, Spanish, German, and Italian and a nuanced understanding of grammar and cultural context. However, like many large models, it can also be used in Chinese after actual testing (there are test cases later).

Of course, the most important thing is the open source attribute of MistralAI——In addition to using cloud service deployment, it can also be directly downloaded and deployed locally. Unlike OpenAI, the leader in closed AI models, MistralAI's large models can be downloaded on code platforms such as "Huahuanglian", but there is no documentation for the Large model yet.


(Source: Company official website)

The company also released its first chat robot LECHAT on Monday, which can choose from the just-released Large model, MistralSmall, which optimizes latency and cost, and a next-generation prototype model called Next. All in all, it provides a window for trial use.

Judging from the short hands-on test, this model has no problem answering simple reasoning questions even if it is used in Chinese.


(Source: LECHAT)

For American history and French history, this model can also answer the questions.


However, in additional tests, this model's understanding of Chinese culture and history can be said to be excellent. ． ． ． It’s terrible, but Mistral itself doesn’t say that this model understands Chinese history. This is also a common problem with foreign open source AI models.



(Source: LECHAT, Wen Xinyiyan)

Microsoft acquires part of the company

Also on Monday, MistralAI and Microsoft announced a new partnership. Judging from the announcement, the cooperation between the two parties mainly includes three aspects: 1. Microsoft provides MistralAI with the computing power required to develop and operate models; 2. MistralAI's large models are put on Microsoft cloud services for deployment by global users; 3. The two parties will explore cooperation to produce specific models for specific customers.


In addition to providing computing power and cloud service customers,This cooperation agreement also allows Microsoft to hold a "small stake" in MistralAI.

It is worth mentioning that the original meaning of the French word Mistral refers to "strong wind blowing from France".The company was founded with the vision that AI does not require so much money. Many companies have wasted a lot of money. Founder and CEO Arthur Mensch bluntly stated that the company wants to be the most capital efficient in the AI ​​world.

Mensch revealed that the Large model released on Monday,The entire development cost is less than 20 million euros. For comparison, Altman mentioned last year that the development cost of GPT-4 could exceed $100 million.

For Microsoft, whether OpenAI, which represents the closed ecosystem, or Mistral or Meta, which represents the open source community, outperforms, the company will be the big winner in the AI ​​track. The AI ​​models of these three companies have been put on Azure's cloud services, and now the "King of Stocks" in the US stock market still holds equity in the first two companies.