China's AI industry map: who is the "capital of large models"

Source: Economic Observer (ID: eeo-com-cn), Authors: Ren Xiaoning, Tao Siyu, Tan Yuqing

Image source: Generated by Unbounded AI tool

A regional race for large models has begun.

"More than 80 large models have been released to the public, of which about half are in Beijing," Jiang Guangzhi, Secretary of the Party Leadership Group and Director of the Beijing Municipal Bureau of Economy and Information Technology, said in a recent interview with the media.

According to incomplete statistics from the Economic Observer, in the new round of generative AI boom, 106 large-scale models have emerged across the country, of which the number of large-scale model companies and research and development companies in Beijing occupy the "C position", as many as 54; Shanghai has accumulated advantages in AI computing power; Shenzhen, Guangzhou, Hangzhou, Chengdu and other cities also have many large-scale model companies.

On April 28 this year, a meeting of the Political Bureau of the CPC Central Committee proposed that it is necessary to attach importance to the development of general artificial intelligence, create an innovative ecology, and pay attention to risk prevention. The first meeting of the 20th Central Financial and Economic Commission held on May 5 pointed out that it is necessary to grasp the wave of new technological revolutions such as artificial intelligence. Afterwards, support policies for large-scale models were successively launched across the country. **Seize a new round of AI highlands, major cities are in action. **

At present, the large models in various places are still in internal testing and are not open to the public. The development of large-scale models is still at the beginning of chaos, and it remains to be observed who will win in the end and become the real "capital of large-scale models".

1. Beijing: 54

Jiang Guangzhi's words are not exaggerated. According to the incomplete statistics of the Economic Observer, as of July 13, there are already 54 large-scale model companies in Beijing.

Zhongguancun is the core gathering place of Beijing large model companies. When a reporter from the Economic Observer visited Wudaokou, there were nearly 10 large-scale start-up companies within a 3-kilometer radius of **, and they were all outstanding among China's large-scale model companies.

Outside the south gate of Tsinghua University in Wudaokou, on the 3rd floor of Sohu Internet Building, it is light years away from the establishment of Wang Huiwen. The company was acquired by Meituan at a price of 2.065 billion yuan just four months after its establishment. Opposite Sohu Building is Zhiyuan Building, where Beijing Zhiyuan Artificial Intelligence Research Institute is located. The dean is Huang Tiejun, director of the Department of Computer Science and Technology of Peking University. This is one of the backbone of the academic school. As early as 2021, Zhiyuan launched China's first ultra-large-scale pre-training model "Enlightenment 2.0", with a parameter volume of 1.75 trillion.

800 meters to the south of Zhiyuan Building is Baichuan Intelligence founded by Wang Xiaochuan. On July 11, Baichuan Intelligence released a general-purpose large language model "Baichuan-13B-Base" with 13 billion parameters.

On Danling Street, 3 kilometers away from Zhiyuan Building, the MiniMax Beijing team works here. This is currently the most highly valued large-scale model start-up company in China. After the latest round of financing in June, it was valued at US$1.2 billion. One kilometer away from Danling Street is the office of Lianyuan Technology, a company founded by Zhou Bowen, the former head of AI at JD.com.

Located near Zhongguancun, there is also Lee Kai-fu's new large-scale model company "01Wanwu". This company was just established on July 3, and the team currently has dozens of core members.

The company of Li Zhifei, the founder of Going Out and Asking, is also in Haidian, and they have launched a large-scale model product "Serial Monkey". There are also large-scale model products launched by major Internet companies nearby. For example, Bytedance located on Zhichun Road released a large-scale model service platform "Volcano Ark" on June 28.

If the scope of Zhongguancun is expanded to a radius of 10 kilometers, there will be more large-scale model companies. 10 kilometers to the north of Zhongguancun, Houchang Village is the gathering place of Internet giants. In March of this year, Baidu released the large-scale model of Wenxin in Baidu Science and Technology Park, which is the first large-scale model released by a major Internet company in China.

The same big factory that released the big model in Beijing, and 360 in Wangjing, released the "Smart Brain Big Model" this year. Jingdong, located in Yizhuang, released the "Yanxi Large Model".

There are also a number of medium-sized companies in Beijing that have also released large-scale models, such as Kunlun Wanwei's "Tiangong", Zhihu's "Zhihaitu", NetEase Youdao's "Ziyue", 4Paradigm's "Shishuo"... …

According to IDC's "2022-2023 China Artificial Intelligence Computing Power Development Evaluation Report", Beijing ranks first in the list of China's artificial intelligence cities in 2022. According to the "AIGC Employment Trend Big Data Report 2023" released by Liepin Research Institute, among the top 10 cities in AIGC's new job distribution in the past year, Beijing has the most jobs, accounting for 22.21%.

**In terms of AI policy support, Beijing is making efforts. **In May, the Beijing Municipal Government issued 3 AI-related documents within 12 days. On May 19, the "Beijing General Artificial Intelligence Industry Innovation Partnership Plan" was issued; on May 30, "Several Measures for Beijing to Promote the Innovation and Development of General Artificial Intelligence" were issued; The implementation plan of the source of artificial intelligence innovation with global influence (2023-2025)".

Jiang Guangzhi also mentioned that the Beijing Municipal Economic and Information Bureau will iteratively launch large-scale model application results on a quarterly basis, forming a number of typical cases of deep integration of artificial intelligence and economic and social development, and is planning to support model partners and models through computing coupons and other forms Observers, the first phase is expected to support no less than 40 million yuan in computing power coupons, subsidized to model partner companies.

2. Shanghai: 12

On July 8 this year, at the closing ceremony of the 2023 World Artificial Intelligence Conference, "Several Measures for Promoting the Innovation and Development of Artificial Intelligence Large-scale Models in Shanghai" was announced, and the "Model" initiative was released to establish the Shanghai Artificial Intelligence Open Source Ecological Industry Cluster. Create an AI "Model City".

Focusing on the four directions of innovation capability, innovation elements, innovation application and innovation environment, Shanghai will focus on building three plans and five major platforms. The three plans are: Large Model Innovation Support Plan, Intelligent Computing Power Acceleration Plan, and Demonstration Application Promotion Plan.

However, compared with the number of large-scale model enterprises in Beijing, Shanghai is slightly inferior. According to incomplete statistics from the Economic Observer, as of July 13, there were 12 large-scale models in Shanghai, namely: Shanghai-based MiniMax, Fudan University’s MOSS, SenseTime’s Ririxin, Daguan Data’s Cao Zhi, Weimob The WAI of WAI, Wuya and Qusao of Transwarp Technology, Leyan of Leyan Technology, Uni-talk of Medical Computing Network, the general large-scale model system for scholars of Shanghai AI Lab, TigerBot of Hubo Technology, and Xiaoi of Zhizhen Intelligence , honey degree of honey degree Wenxiu.

The leaders of the Shanghai model are SenseTime and Fudan University. In July this year, the large model of SenseTime was updated to version 2.0, and a small model was also launched on the basis of the large model. Fudan University's MOSS was launched in February this year. Qiu Xipeng, the person in charge of the MOSS system, said that MOSS has surpassed ChatGPT in terms of Chinese ability.

Although the number of large-scale models is not as large as that of Beijing, **Shanghai is far ahead in terms of computing power and chips that every large-scale model company cannot do without. **

In the first half of this year, Nvidia's stock price soared because their GPUs can provide computing power for large models. Due to the limited shipments of GPUs, the market is in short supply. ** In Shanghai, there are also a number of domestic GPU companies. **

Minhang District and Shanghai Pilot Free Trade Zone are important centers for Shanghai chip companies. On Chenxing Highway, there are leading domestic chip companies that provide general-purpose GPU chips for large-scale calculations, such as Biren Technology and Tianshu Zhixin. In the Shanghai Pilot Free Trade Zone, there are many semiconductor companies with different nodes and types of research and development from chip design to production, such as Xinlu Technology, Riguan Xinshi, and Shitong Microelectronics.

Zhangjiang is also a gathering place for Shanghai chip companies. At this year's World Artificial Intelligence Conference, Zhangjiang's chip companies released new chip products, including the "Xisi N100 Artificial Intelligence Inference GPU" integrated by Muxi, the "Hanbo second-generation full-featured GPU chip" of Hanbo Semiconductor, and the Yuan Technology's "Artificial Intelligence Cloud Training Chip Deep Thought 2.0" and so on.

At the same time, there are many foreign-funded chip companies in Shanghai. In Zizhu Science Park, Minhang District, there is a Chinese company with semiconductor chip giant Intel, and Nvidia, which defines GPU, has its Chinese company in Shanghai.

In addition to large models, Shanghai is also the birthplace of the last wave of "AI Four Tigers" SenseTime and Yitu Technology that deeply cultivated computer vision. In 2019, Shanghai Pudong New Area became the country's first artificial intelligence pilot area. Now, 13 colleges and universities including Shanghai Jiaotong University, Fudan University, Tongji University, and Shanghai University, and 5 scientific research institutes have established artificial intelligence-related disciplines.

In terms of AI support policies, Shanghai is running fast across the country. In September last year, Shanghai promulgated the country's first AI provincial-level regulation "Shanghai Municipal Regulations on Promoting the Development of Artificial Intelligence Industry", which mentioned that Shanghai promotes the transaction and circulation of algorithm models, and the Shanghai Municipal Economic Information Department, together with relevant departments, guides industry organizations to formulate artificial intelligence algorithms The model recommendation catalog reduces the risk of algorithmic model transactions and improves the efficiency of algorithmic model transactions. The "Regulations" pointed out that Shanghai supports relevant entities to carry out high-efficiency smart chip design innovation based on advanced architecture, develop cloud chips and cloud smart servers, deploy brain-like chips, strengthen software and hardware collaborative adaptation, and support joint construction of research and development testing and application platforms , strengthen technical collaboration and chip architecture interconnection, enhance the competitiveness of smart chip products and technologies, and cultivate a smart chip application ecosystem.

3. Shenzhen: 11

In the administrative service hall of Shenzhen Longgang Political and Digital Bureau, the large-scale model has been applied to the government consultation service system, and the large-scale model named "Tianshu" of Yuntian Lifei Technology Co., Ltd. is behind it. This AI company landed on the Science and Technology Innovation Board on April 4 this year. It is "Shenzhen's first artificial intelligence stock", with a current market value of 22.9 billion yuan.

Zhuiyi Technology, another natural language processing technology company in Shenzhen, also released its self-developed large-scale model "Bowen Bowen" this year, which landed in vertical scenarios such as finance and government affairs.

In addition, Shenzhen also has large-scale model research capabilities. In Futian, Shenzhen, the IDEA Research Institute led by Shen Xiangyang, the former global executive vice president of Microsoft and the head of AI and technology R&D system, launched the open source general model "Jiang Ziya" series and the 99th general model "Jiang Ziya general model" on the Fengshen list v1". Jiang Ziya's large model has 13 billion parameters and can be pre-trained at the level of 100 billion tokens. At the same time, the Chinese University of Hong Kong (Shenzhen) and Shenzhen Institute of Big Data trained and open sourced a large medical model - HuatuoGPT (HuatuoGPT), so that the language model has the ability to diagnose and provide useful information like a doctor.

"Beijing does things from 0 to 1, Shanghai does things from 1 to 10, and Shenzhen does things from 10 to 100." A researcher who focuses on Shenzhen's artificial intelligence companies described Shenzhen's AI field It is different from Beijing and Shanghai. In his view, AI companies in Shenzhen are good at top-level AI applications, and it may take half a year to a year before Shenzhen companies can show their talents.

Not long ago, the above-mentioned people did research on local large-scale model companies. The latest data shows that there are 11 large-scale model companies and institutions in Shenzhen, including both general-purpose large-scale models and industrial large-scale models.

Of course, the undisputed leading company of AI in Shenzhen is still Huawei. On the afternoon of July 7th, Huawei released the Pangu Large Model 3.0, and Huawei also provided computing power. On July 6th, Huawei's computing power solution Shengteng AI cluster was fully upgraded, and the cluster scale was expanded from the initial 4,000-card cluster to 16,000-card. , the industry's first Wanka AI cluster.

On May 31, Shenzhen issued the "Action Plan for Accelerating the High-quality Development and High-level Application of Artificial Intelligence in Shenzhen (2023-2024)", simultaneously releasing the first batch of "city + AI" application scenario lists, and coordinating the establishment of a scale of 100 billion yuan With the most sufficient computing power, the greatest policy support, the best industrial ecology, the best talent environment, and the richest scenario applications, we will build a new generation of national artificial intelligence innovation and development pilot areas and national artificial intelligence Innovative application pilot zone.

The above-mentioned researchers told the Economic Observer that due to the relatively high threshold for starting a large-scale model, top talents are required, and the focus is on basic research. In the early stage of large-scale model development, a city with many universities like Beijing has an advantage. **Shenzhen has a well-developed manufacturing industry and will focus on application and industrialization in the future. **

On July 4, the "Shenzhen Artificial Intelligence Industry Development White Paper (2023)" released by the research team of Shenzhen Artificial Intelligence Industry Association showed that by 2022, the scale of Shenzhen's artificial intelligence industry will reach 248.8 billion yuan. **

Among them, the scale of Shenzhen's artificial intelligence core industry reached 30.8 billion yuan, a year-on-year increase of 52.48%; in terms of the number of enterprises, in 2022, the number of artificial intelligence enterprises in Shenzhen reached 1,920; in terms of industrial chain distribution, in 2022, Shenzhen's artificial intelligence basic layer Accounting for only 8.13% of the total number of enterprises, the technical layer accounts for 17.29% of the total number of enterprises, and the application layer accounts for 74.58% of the total number of enterprises; in terms of regional distribution, in 2022, Shenzhen's artificial intelligence enterprises will mainly be concentrated in Nanshan and Bao'an districts. The total proportion is as high as 58.1%. Among them, the number of Nanshan District accounted for 36.8%, Baoan District accounted for 21.3%.

4. Hangzhou: 8

According to the incomplete statistics of the Economic Observer, there are 8 large-scale models in Hangzhou. The largest is the Tongyi large-scale model family of Alibaba Cloud. Zhenyi of Ant Group, Baiye Lingxi of H3C, Tas of real intelligence, and Yuyan of Netease Fuxi.

Hangzhou’s AI leader is Alibaba. At present, Alibaba Cloud’s Tongyi large-scale model family already has 3 large-scale models, including Tongyi Qianwen, Tongyi Tingwu and Tongyi Wanxiang. Ant Group, a subsidiary of Alibaba, also released a large model of Zhenyi.

Another AI force in Hangzhou comes from NetEase. At the 2023 World Artificial Intelligence Conference, NetEase Fuxi made its debut with a variety of AI large-scale model products, including "Yuyan" and "Danqing".

On the edge of West Lake in Zhejiang, there is also a large-scale model start-up company, Xihu Xinchen. This company cooperates with scientific research institutions such as West Lake University, integrates the strength of academic schools, and releases a large model of West Lake. Since the beginning of this year, Xihu Xinchen has received investment from Baidu and Tom Cat. Its latest round of financing occurred in July, and it completed a strategic financing of over 100 million yuan.

According to the series "Evaluation Report on the Development of China's Artificial Intelligence Computing Power" jointly released by IDC and Inspur Information, in the past five years, Hangzhou has steadily ranked among the top three cities in the top ten computing power cities, competing with Beijing and Shenzhen, showing strong competition force. **

In terms of computing power carrier, Hangzhou has Aliyun City Brain, Hikvision Video Sensing National New Generation Artificial Intelligence Open Innovation Platform, Brain and Brain-Computer Fusion Frontier Science Center, Artificial Intelligence Collaborative Innovation Center, and National Data Intelligence Technology Innovation Center and other national innovation platforms.

At the same time, technology companies in Hangzhou have also found a ready-made testing ground: the e-commerce industry. As the capital of e-commerce, the need for frequently updated content production has given rise to the market demand for AIGC from e-commerce, which has become a pillar project for AIGC's marketization in Hangzhou. At present, the AIGC service of Hangzhou technology companies has been able to realize the rapid construction of e-commerce scenarios. In addition to marketing texts, product introductions and picture materials required by merchants can be automatically generated through AIGC. The AI painting that was popular last year, the AI content generation that some new media are already using, etc., Hangzhou enterprises have also launched related products.

In terms of policies, at the end of last year, Hangzhou released the "List of Artificial Intelligence Application Scenarios for the Third Batch of Key Constructions in Hangzhou in 2022", covering smart cities, smart manufacturing, smart medical care, smart transportation, smart finance and other fields, with a total of 46 key constructions Artificial intelligence application scenarios were selected. In June of this year, the Hangzhou Economic and Information Bureau drafted the "Implementation Opinions on Accelerating the Innovation and Development of the Artificial Intelligence Industry (Draft for Comment)", which is now open to the public for comments.

At present, Zhijiang Laboratory, Alibaba Dharma Institute, Zhejiang University and other scientific research institutions are increasing their research in the field of AIGC.

5. Guangzhou: 4

According to the incomplete statistics of the Economic Observer, there are four large models in Guangzhou, namely Congrong from Cloudwalk Technology, Bian Que and Lingxin from the Future Technology College of South China University of Technology, SocialGPT from Data Storytelling, and Darwin from Sailingli Technology.

In July 2021, the Guangzhou Municipal Bureau of Science and Technology released the "Three-Year Action Plan for the High-Quality Development of the Guangzhou Artificial Intelligence Industry Chain (2021-2023) (Draft for Comments)", proposing that Guangzhou will focus on creating an internationally influential The AI Innovation Contest, the construction of 10 AI industrial parks, the creation of 100 typical AI application scenarios, and the cultivation of over 1,000 AI companies.

In September last year, Guangzhou issued the "Guangzhou City Action Plan for Promoting the Integrated Development of the Innovation Chain and Industrial Chain (2022-2025)", pointing out six key actions.

6. Chengdu: 3

According to incomplete statistics, there are three large-scale models released in Chengdu: the Xiao model XPT jointly released by Xiaoduo Technology and the National Supercomputing Chengdu Center, the medGPT released by Medlink Technology, and the workChat digital human of Mingtu Technology.

In May of this year, Xiao Model (XPT) was released, which is a large vertical model for the e-commerce industry developed based on large language model technology. On May 25, MedLink released the first domestic medical large language model - MedGPT, which is committed to giving full play to the value of actual diagnosis and treatment in real medical scenarios, and realizing the full-process intelligent diagnosis and treatment capabilities from disease prevention, diagnosis, treatment, and rehabilitation. On May 31, Mingtu Technology officially released the Mingtu workChat digital human, whose functions include knowledge question and answer, AI writing, knowledge collection, auxiliary tools, etc.

In May of this year, the "2023 Chengdu City Big Data and Artificial Intelligence (Including Vehicle Intelligent Control System) Industrial Circle Building and Strong Chain Work Points" was issued, with 17 detailed measures to give Chengdu big data and artificial intelligence (including vehicle intelligent control system) ) Industrial chain development "road map".

The goal of Chengdu is to strive to cultivate 5 national-level artificial intelligence specialized and new "little giant" enterprises, 1 listed company, and 10 listed enterprises this year, and the scale of the artificial intelligence industry will exceed 70 billion yuan.

Chengdu City proposes to encourage universities, scientific research institutions, and leading enterprises to combine computing power centers to focus on incubating original AI large models in key industrial fields, and to give construction parties a computing power cost subsidy of up to 5 million yuan per year to reduce corporate innovation costs. A total of no more than 10 million yuan of computing power coupons will be issued every year to support computing power intermediary service agencies, technology-based small and medium-sized enterprises and makers, scientific research institutions, universities, etc. to use the computing power of the National Supercomputing Chengdu Center and Chengdu Intelligent Computing Center resource.

VII. Nanjing: 3

Nanjing, which is located in the Yangtze River Delta region, has also been making efforts to develop the AI industry in recent years. According to incomplete statistics, there are already three large-scale model companies in Nanjing, namely KidsGPT from Kidswant, Yuanchengxiang from Sophon Engine, and Yandi from Silicon-based Intelligence.

Kidswant’s KidsGPT was launched during the Children’s Play Festival, specializing in the vertical field of mothers, babies and children. The Yandi large model is a multi-modal industry model, which can be used by enterprises to generate short videos, live broadcasts, movies/tv series, and even digital immortality. Sophon Engine is a multi-modal large-scale model start-up company that has completed tens of millions of angel round financing. The core team behind the company is the research group led by Lu Zhiwu, a professor at the Hillhouse School of Artificial Intelligence at Renmin University of China. Its large-scale model product, the meta-multiplication image ChatImg, has a parameter of about 15 billion.

In December 2022, Nanjing issued the "Implementation Plan for the Construction of Nanjing National Artificial Intelligence Innovation and Application Pilot Zone", proposing that by 2025, Nanjing will build 5 artificial intelligence industrial parks with high standards and strive to create 1 or 2 artificial intelligence centers above the provincial level. Industrial clusters, cultivating 100 key artificial intelligence enterprises, 30 provincial-level or above specialized small giant enterprises and individual champion enterprises, and the city's artificial intelligence core industry revenue exceeding 50 billion yuan. **

In the future, Nanjing Artificial Intelligence Computing Center will continue to provide enterprise customers with easy-to-use services and products in addition to computing power support, focusing on data, computing power, algorithms and other core capabilities of the artificial intelligence industry to empower enterprises and promote the artificial intelligence industry Upstream and downstream enterprises gather in Nanjing. The center plans to complete the model adaptation verification of core partners in the near future, which will provide a strong basic model support for the large language model of the local business incubation industry, and finally create a benchmark demonstration in Jiangsu for the transfer of large-scale industry models to business. **

8. Others: 11

Other cities that released large models include Tianjin, Hefei, Harbin, Shenyang, Xi'an, Changsha, Mianyang, and Xiamen.

There are already two large-scale model companies in Tianjin: the Tianhe Tianyuan model released by the National Supercomputing Tianjin Center, and the Haihe Diting model jointly released by Huiyan Technology and Tianjin University.

During the World Artificial Intelligence Conference, the National Supercomputing Tianjin Center released the Tianhe Ten Billion Level Intelligent Computing Open Innovation Platform and the domestic Chinese large model - Tianhe Tianyuan. The National Supercomputing Tianjin Center has collected and sorted out webpage data, various open source training data, Chinese novel data, ancient Chinese data, encyclopedia data, news data, and various data sets and training data in professional fields such as traditional Chinese medicine, medicine, consultation, law, etc. The total number of tokens reached 350B.

The representative of the AI company in Hefei, Anhui is HKUST Xunfei. In May of this year, HKUST Xunfei released the Xinghuo Cognitive Model. Liu Qingfeng, chairman of HKUST Xunfei, said that the current Xunfei Xinghuo cognitive model has surpassed ChatGPT in the three major capabilities of text generation, knowledge question and answer, and mathematical ability.

In the Hefei High-tech Zone, an industrial park centering on the field of intelligent voice and artificial intelligence has also flourished, known as the "Voice Valley of China". More than 1,000 enterprises have settled here, employing more than 50,000 people, forming an industrial cluster worth 100 billion yuan.