Meituan sells large models again and exclusively invests in “Zhipu AI”
The big AI model startup with hot money has come to an end, and a key investment around the head AI startup company has come to an end.
36 Krypton learned from multiple independent sources that the AI large-scale model company Zhipu Huazhang (hereinafter referred to as “Zhipu”) has completed the B-2 round of financing a few months ago, with an amount of hundreds of millions of RMB, which was exclusively invested by Meituan Zhantou. The post-investment valuation was nearly $500 million. It is understood that Zhipu’s subsequent rounds of financing have also been completed one after another.
Regarding the above financing information, the Zhipu official has not yet expressed its position.
This is also one of the few recent financings with an amount exceeding US$100 million. Most of the AI start-up companies that have received financing before are at the level of one million or tens of millions of dollars. The few that can produce self-developed models are still one step away from the billion-dollar club.
An investor told 36 Krypton that a start-up company that can earn hundreds of millions of dollars needs to have both a top-level team and self-developed large-scale technology, as well as successful landing experience-and such companies can be counted on one hand in China.
In the field of AI large-scale model research, the Zhipu team has always attracted attention in the industry. It was established in 2019 and transformed from the technical achievements of the Knowledge Engineering Laboratory (KEG) of Tsinghua University.
Born in the Natural Language Laboratory of Tsinghua University, Zhipu’s core team lineup can be described as luxurious: Zhang Peng, the designer and developer of XLORE, the first Chinese-English balanced cross-language knowledge graph system in China, is the CEO, and most of the team’s research members are from the Department of Computer Science of Tsinghua University.
Data, computing power, and talents are the necessary tickets for large-scale entrepreneurs. As early as 2021, Zhiyuan basically gathered all the elements.
An insider told 36 Krypton that thanks to the open source intelligence analysis project started nearly 20 years ago, the team not only launched AMiner, a big data mining and service system platform for scientific and technological intelligence in 2006, but also accumulated rich Chinese-English bilingual data and first-line engineering experience for the development of large models. And large-scale computing power is solved by renting supercomputing platforms and cooperating with research institutions such as Zhiyuan.
In August 2022, Zhipu and KEG jointly launched the 100 billion-level Chinese-English pre-training language model GLM-130B. According to observations, apart from Internet companies such as Alibaba, Zhipu was the first start-up to try a large language model with a scale of 100 billion at that time.
The performance of the GLM-130 B was also recognized. A study by the Center for Fundamental Model Research at Stanford University showed that GLM-130 B can be close to or equal to the large model GPT-3 175 B released by OpenAI in terms of accuracy and maliciousness indicators. This model, which established the position in the smart spectrum industry, has been applied for by 266 institutions in 41 countries around the world. According to 36 Krypton, the institutions that applied for the use include Tencent, ByteDance and other leading Internet companies.
In March of this year, Zhipu open-sourced the ChatGLM-6 B, a lightweight model of Chinese-English bilingual dialogue with 6.2 billion parameters, which supports users to use inference on a single consumer-grade graphics card. At the same time, ChatGLM, a language generation application based on the GLM-130 B base, was also launched at the same time, and internal testing was started.
Even though there are still common problems of large models such as mathematical processing and weak logical reasoning ability, ChatGLM-6 B has made up for the time-sensitive shortcomings that ChatGPT has not yet solved. A person familiar with the matter introduced to 36 Krypton that ChatGLM-6 B knew that the current CEO of Twitter is Elon Musk due to the technology accumulated by the previous team in business cooperation and research.
With the blessing of multiple resources and technical capabilities, Zhipu is particularly popular among the new AI forces.
In September 2022, Zhipu completed the B-round financing of 100 million yuan, which was jointly led by Legend Capital and Qiming Venture Capital. The entry of this round of Internet capital means that the development of large models has quickly entered the stage of application and implementation from research exploration.
A person familiar with the matter told 36 Krypton that Zhipu’s algorithm has already performed relatively well in the optimization of Meituan’s advertising algorithm.
Meituan has been looking for opportunities to combine AI and business. As early as February 2018, Meituan Dianping established the Al Platform Department-NLP (Natural Language Processing) Center to provide text semantic understanding services for various businesses including intelligent customer service, scene-based store search and search. Wang Zhongyuan, the former Facebook (now Meta) Research Scientist and Microsoft Asia Research Institute chief researcher, and now the vice president of Kuaishou Technology, is the leader of the establishment of the NLP Center.
In the organizational structure adjustment in mid-2020, Meituan Dianping search and NLP departments were incorporated into the newly established AI platform, led by Xia Huaxia, chief scientist of Meituan, who was a senior engineer at Google and chief architect at Baidu.
Since the explosion of ChatGPT, it is not uncommon for Internet companies to rely on investment in start-ups to build an AI ecosystem. For example, Zhihu invested in Mianbi Smart and jointly developed a large-scale model “Zhihaitu AI”. Holding hands with Zhipu, Meituan has found an ally with more mature technology and business.