The European version of OpenAI exploded, and Llama created its own portal

Ming Min from the Au Fei Temple

Qubits | Official account QbitAI

The latest valuation of the "European version of OpenAI" is approaching $2 billion!

Paris-based model startup Mistral AI's latest funding round of $487 million.

Investors include Anderson Horowitz, Nvidia, Salesforce, etc.

Among them, Anderson Horowitz led the investment of about 200 million euros, and Nvidia and Salesforce plan to invest another 120 million euros through the conversion of bonds.

It is understood that Mistral AI was founded by former DeepMind and former Meta scientists. The 2 main creators are the core contributors of Llama.

In May this year, they started a business after leaving, which quickly attracted attention in the circle. The seed round raised $113 million, valuing the company at $260 million.

In September, the open-source large model Mistral-7B was released, which fully benchmarked against Llama2.

Now, with the latest round of funding entering the final stage, it means that Mistral AI has successfully entered the ranks of unicorns in just 6 months.

According to reports, the new funds will be used to expand the team and accelerate research and development.

Llama was founded by the main creator of Llama

Founded in May this year and headquartered in Paris, Mistral AI's mission is to "build the next generation of AI systems that solve real-world problems."

Due to its ambitious AI vision, the fact that its creators are all top talents in the AI field, and the rapid release of self-developed large models, Mistral AI is regarded by the industry as the "European version of OpenAI".

The company's three main creators are:

CEO：阿瑟•门奇(Arthur Mensch)

Chief Scientist: Guillaume Lample

CTO:蒂莫西•拉克鲁瓦(Timothée Lacroix)

Among them, the chief scientist and CTO are the core creators of Llama.

Chief scientist Guillaume Lampel was a research scientist at Meta AI and led the development of Llama.

He received his Ph.D. from Carnegie Mellon University under the supervision of Chris Dyer, where his research interests were in the areas of NLP and machine learning.

According to Google Scholar, Lampel's paper has been cited 19,000+ times.

CTO Timothy Lacroix spent eight years at Meta AI and is a major contributor to PyTorch-BigGraph.

CEO Arthur Mench, 30, was a researcher at DeepMind in Paris before starting his business and started his business after resigning in 2023.

During his time at DeepMind, Menchi was a major contributor to projects such as Retro, Flamingo, and Chinchilla, and accumulated rich experience in the development and optimization of large language models.

It is understood that the three main creators met when they were in school, when they were all studying in the field of AI, and CEO Menqi said that they began to discuss the idea of AI entrepreneurship in 2022.

In June this year, just four weeks after its launch, Mistral AI won a $113 million seed round, led by top venture capital Lightspeed, and reached a valuation of $260 million.

At that time, the team said that they would work on the open source of large models.

In September, Mistral AI officially released and open-sourced its first large model, Mistral-7B, and comprehensively benchmarked against Llama2.

According to the paper, Mistral-7B surpassed Llama 2 (13B) in all benchmarks and Llama 2 (34B) in reasoning, math, and code generation tasks.

Mistral-7B adopts the same group query attention mechanism in Llama2, which can accelerate inference, and combined with the sliding window attention mechanism, it can effectively process arbitrary length sequences and reduce the inference cost.

They also provide a fine-tuned model, Mistral 7B – Instruct, that also outperforms the Llama2 (13B) chat model in both human and automated benchmarks.

Mistral-7B shows that LLMs may be able to compress knowledge more than we previously expected. If you want to get better performance with a small model, there's a lot of room to explore.

It is understood that in order to avoid copyright issues, Mistral uses public data to train large models.

Become a rising star of European AI

With the release of its own large model, Mistral has also officially competed with OpenAI, Google, etc.

CEO Mench once said in an interview that although the company is young and the team is small, "we compete with everyone".

Due to the gathering of many top talents in the field of large models, Mistral has attracted much attention in the industry since its inception, especially in Europe.

Its seed round attracted many established European venture capitalists, such as Redpoint and Index Ventures in France, La Famiglia and Headline in Germany, as well as venture capital from Italy, Belgium and the United Kingdom.

At the inaugural AI Security Summit recently held in London, Mistral, as the only European company, attended the second day of a small conference with tech giants.

Moreover, Mistral did not disappoint investors.

It is said that the release of its first large-scale model was expected to be launched at the end of the year, but it was released in September, three months ahead of schedule.

With a top team lineup and a self-developed model, it is not difficult to understand the rapid growth of Mistral's valuation.

The latest round of funding was mainly contributed by Silicon Valley.

The new funds will be used to expand the team, accelerate R&D, and establish new offices in Europe and the United States.

According to reports, as part of the deal, Mistral's three co-founders will each sell more than £1 million of their shares.

This huge round of funding also confirms once again that investors are scrambling to find the "next OpenAI", and many large investors are willing to buy shares at high prices in the early stages of startups, which is one of the reasons why startups soar in valuations in the early stage.

Reference Links:

[1]https://www.businessinsider.com/mistral-in-talks-to-raise-funding-at-2-billion-valuation-2023-11

[2]https://techstartups.com/2023/12/04/openai-rival-mistral-ai-nears-2-billion-valuation-with-backing-from-andreessen-horowitz-nvidia-and-salesforce/

[3]https://www.ft.com/content/387eeeab-1f95-4e3b-9217-6f69aeeb5399