AI is surging, and the king of new topics is Mistral AI, whose newly launched large model scores second only to OpenAI's GPT-4.
This AI start-up from France, founded only 9 months ago, the founding members are all around 30 years old, attracted a huge amount of funds in a short period of time, quickly broke out of the encirclement of giants, and now hand in hand with Microsoft, there is a "European AI model leader", "the next OpenAI" posture.
After the release of the new flagship model Mistral Large in the early morning of February 26, Beijing time, it immediately set off a huge wave in the social circle, and the server has been crowded with users from all over the world.
However, Mistral Large, while performing well, is a closed-source route that is neither open nor free, which is contrary to the company's vision when it was first founded. Some netizens pointed out that Mistral AI even deleted the content on its official website about the obligations of the open source community.
An AI large model entrepreneur told the 21st Century Business Herald reporter that the company will eventually take the path of commercialization, "Mistral will open source the small model, and the large model will be closed-source, and the good effect will be used to make money, which is also normal." The person lamented that in the AI era, capital and computing power are absolute bargaining chips.
After all, the business world has to consider benefits, not to mention that in the field of extremely money-burning AI models, the "money ability" obtained by financing alone cannot support the company to go far.
According to incomplete statistics, since its establishment, Mistral AI's total financing has exceeded $500 million, and its current valuation is over $2 billion. The giants are targeting the opportunity to get a slice of the mistral stake in order to get a head start in the future AI wars.
"Leaderboard".
Mistral AI's newly released Mistral Large model is described by the company as "achieving top-level inference capabilities and can be used for complex multilingual inference tasks, including text understanding, conversion, and generation." At the same time, the company simultaneously launched a chat assistant called Le Chat, through which users can call up its latest large model.
In a number of officially announced tests, the Mistral Large has achieved excellent results in commonly used benchmarks, making it the world's second-ranked model, behind OpenAI's GPT-4, better than Anthropic's Claude 2 and Google's Gemini Pro, and far ahead of GPT-35 and meta's llama2.
Source: Mistral AI
According to Mistral AI, Mistral Large has four features and benefits. First, Mistral Large is able to apply English, French, Spanish, German, and Italian fluently like a native language, with a nuanced understanding of grammar and cultural context. This is undoubtedly a comfort zone for Mistral AI, which was born in Europe.
Second, Mistral Large is able to handle the contextual content of 32k tokens, allowing it to accurately recall information from huge documents. In other words, the model is not afraid of pressure when it comes to processing long documents.
Third, Mistral Large is exceptionally precise in executing specific instructions, which allows developers to tailor content moderation policies to their needs. It is reported that Mistral AI used it to conduct a system-level review of Le Chat.
Finally, mistral large supports function calls. Combined with the output content restriction mode implemented by Mistral AI on La Plateforme, it is possible to modernize the development of applications and the technology stack.
GPT-4 below, 10,000 models", so that Mistral Large attracted many users to try to see its true face. Before the day on February 27, Beijing time, some netizens did post their evaluation feelings about mistral large, and tested the typical "chicken and rabbit in the same cage" question in Chinese, and mistral large gave the correct answer.
However, unlike Mistral AI's previous "street bombing" model, Mistral Large will not be open source, but will be available through specific channels, including Mistral AI's own API, or Microsoft's Azure AI Studio and Azure Machine Learning.
Previously, Mistral AI's official website had explained the company's vision in detail, emphasizing the provision of open-source models as an alternative to counter the monopoly of AI oligarchs. Founder Arthur Mensch has also publicly expressed criticism of closed-source large models.
However, after the release of the new model, some netizens found that the official website of Mistral AI removed the elaboration related to open source. In this regard, foreign netizens expressed considerable dissatisfaction, questioning whether Mistral AI "went against the original intention" and would follow in the footsteps of AI oligarchs.
Commercialization is the company's ultimate proposition, and the path taken by OpenAI may not be avoided by Mistral AI. Some domestic AI practitioners have analyzed it like this.
Arthur Mensch is also reportedly considering how to strike a balance between business models and open source values.
Rapid growth
As an AI upstart, Mistral AI has only been established for 9 months, but it is already a leading AI unicorn company in Europe.
In May 2023, 30-year-old Frenchman Arthur Mensch left Google Deepmind to join Timothée Lacroix, 32, and Guillaume Lample, 33, to found Mistral AI, which originally worked at Meta Platforms' AI lab in Paris.
It is reported that the three founders met during their study period, and subsequently performed well in the large model research and development teams of Google and Meta respectively. Among them, Arthur Mensch joined Google in 2020 and spent three years at the company before founding Mistral AI.
The founding team of Mistral AI believed that small teams surpassed the big companies in Silicon Valley in terms of flexibility, and they chose the open source route, and four months after its founding, they launched the open source model Mistral-7B and successfully challenged Llama 2, which made Mistral famous.
Mensch has revealed that their goal is to be the most capital-efficient company in the AI space. The development cost of the just-launched Mistral Large model may not exceed $22 million. Compared to the cost of OpenAI to train GPT-4, it is only one-fifth.
Why can a small entrepreneurial team make a large model comparable to GPT-4 in the most money-burning field?
Some AI large model practitioners told the 21st Century Business Herald that the fundamental reason is also "abundant funds, money means computing power", although the number of Mistral AI team is small, but there is no shortage of funds, and talents are also gathered.
According to incomplete statistics, when Mistral AI was established only one month ago, it received 10.5 billion euros (about 1.)$1.3 billion) led by Lightspeed Venture Partners, with participation from investors including former Google CEO Eric Schmidt, French billionaire X**ier Niel, and French advertising giant Jcdecaux.
In December 2023, Mistral AI received another round of financing, with a total financing of about 4$1.5 billion. The round was led by Andreessen Horowitz (aka A16Z) and Lightspeed Venture Partners, and Nvidia and Salesforce are also reported to have committed to investing in convertible notes1200 million euros.
Just this week, Microsoft said it would add a new model of Mistral AI to Azure's cloud service options for developers to choose from. According to reports, Microsoft will hold a small stake in Mistral under a multi-year cooperation agreement between the two parties, but Microsoft has not confirmed this statement.
Industry estimates put Mistral AI at a current valuation of about $2 billion, and its shareholders include not only some well-known venture capital firms, but also several tech giants.
Some analysts believe that the rise of Mistral AI has its soil. Compared with Silicon Valley's aggressiveness towards AI, the European market is more cautious, and Mistral AI, as a local AI company in France, will be more tolerant and easier to open the European market than OpenAI.
Microsoft's AI landscape
It is known that behind OpenAI stands Microsoft, which has invested more than $13 billion in it.
Now, Microsoft has announced a high-profile partnership with Mistral AI for many years, and the two sides will also carry out in-depth cooperation in core areas, and it is reported that Microsoft has also joined the list of shareholders of Mistral AI, and there are inevitably voices saying that "Mistral AI is the next OpenAI".
According to Microsoft, Mistral AI can provide its large model on Microsoft's Azure cloud computing platform, and the last company to win this award is OpenAI.
According to the announcement, Microsoft's cooperation with Mistral AI is mainly focused on three core areas, including Microsoft's support for Mistral AI through Azure AI supercomputing infrastructure to support the AI training and inference work of Mistral AI's flagship model; Facilitate the go-to-market pace of Mistral AI by providing customers with large Mistral AI models (MaaS) through Azure AI Studio and the Model as a Service (MaaS) in the Azure Machine Learning Model Catalog; In addition, Microsoft and Mistral AI will collaborate to explore training purpose-specific models for specific customers.
This partnership with Microsoft gives Mistral AI access to Azure's cutting-edge AI infrastructure to accelerate the development and deployment of its next-generation large language models (LLMs) and provides Mistral AI with the opportunity to unlock new business opportunities, expand into global markets, and facilitate ongoing research collaborations.
As for why Microsoft is betting on Mistral AI, from a business logic point of view, this is a normal move for Microsoft to expand its AI territory, through Mistral AI, Microsoft will likely successfully open the European market and find more increments.
Although the Mistral Large model released by Mistral AI this time is not open source, it does not mean that the company will permanently abandon the open source route. For Microsoft, getting involved in both OpenAI and Mistral AI means that it has a layout on both closed-source and open-source routes.
In fact, Microsoft CEO Satya Nadella is optimistic about open source technology, and Microsoft's cooperation with GitHub is based on the pursuit of providing open source tools for developers.
In addition, OpenAI's "palace fighting drama" that broke out last year may also make Microsoft feel uneasy. At that time, as the largest investor, Microsoft could only learn about the OpenAI team's decision to dismiss its CEO Sam Altman a few hours in advance, and OpenAI's shareholding structure determined that Microsoft could not have a say in its major decisions.
Investing in Mistralai can reduce Microsoft's reliance on OpenAI to a certain extent, diversify risks while increasing the insurance coefficient.
Microsoft also said at MWC 2024 that investing in Mistral will help Microsoft continue to innovate, "Innovation and competition will require a wide range of similar support for proprietary and open-source AI models, large and small." ”
It's worth noting that Microsoft has been making a lot of moves in the AI field lately, and is even working on a new network card to improve the performance of its Maia AI server chips. Microsoft, which is on the road of self-developed AI hardware, has also been interpreted by the outside world as trying to reduce its dependence on NVIDIA.
As the person in charge of a leading company investing in the field of AI in China said, the explosion of AI has filled the eyes of investment institutions, and there will always be bubbles.