With GPT skyrocketing and AI investment surging, the entire venture capital market is heated due to artificial intelligence.

Scarce companies turn into unicorns in minutes, new challengers join the race, the AI competition is intensifying, and the "battle of the gods" of China's "large models" has begun.

In February this year, Meituan's co-founder Wang Huiwen entered the "large model" scene with a high-profile "AI Heroes List." This was followed by a wave of heavyweights like former Sogou CEO Wang Xiaochuan, former JD.com AI head Zhou Bowei, and Alibaba's technical vice president Jia Yangqing, dubbed "the number one Chinese in the AI framework field," rushing to join the AI startup ranks.

There are also a multitude of startups with impressive backgrounds emerging. For instance, Tsinghua Professor Tang Jie's ZhiPu AI, Professor Huang Minlie's Lingxin Intelligence, Yang Zhilin's Loop Intelligence... all have been given industry-disrupting hopes by leading venture capitalists, and the startup train is accelerating once again.

The tech media outlet The Information has compiled a list of the top 5 AI startups in China. Who will be China's OpenAI?

1. MiniMax

Funding information: Angel Round

Established: November 3, 2021

Location: Shanghai

Affiliated company: Dream Name (Shanghai) Technology Co., Ltd.

This AI company, established by Yan Junjie, former vice president of SenseTime and head of general intelligence technology, has built a foundational model architecture for text-to-visual, text-to-speech, and text-to-text modalities. It is the startup with the highest valuation in the current large AI model startup wave.

On June 1, according to Reuters, MiniMax completed a new round of financing exceeding $250 million, and the company is currently valued at over $1.2 billion. In this round of financing, an entity related to Tencent participated, contributing perhaps $40 million.

Before this, MiniMax had completed two rounds of financing, with investors including miHoYo, IDG Capital, Gaorong Capital, Yunqi Capital, Mingshi Capital, and others. Yunqi Capital confirmed in a post in April that it invested in MiniMax in 2021, being the only early-stage investor in the angel round.

On February 16 this year, MiniMax revealed at a small media communication meeting in Beijing that it had over a hundred team members. The company's core technical research and development members come from world-renowned universities and top tech companies. They have world-leading experience in natural language processing, voice, computer vision, computer graphics, etc. One third of the team members have PhD degrees from world-leading tech labs.

MiniMax has developed three foundational models from the ground up: text-to-visual, text-to-audio, and text-to-text.

2. Langboat Technology

Funding information: Pre-A+ Round

Established: June 10, 2021

Location: Beijing

Affiliated company: Beijing Langboat Technology Co., Ltd.

Langboat Technology's founder Zhou Ming decided to start a business when the temperature of the domestic AI market hit rock bottom.

At the end of 2020, Zhou Ming considered resigning from his position as vice president of Microsoft Research Asia. Many friends advised him not to, but he was determined to start a large model business, believing that "large models will become some kind of infrastructure in the future."

Zhou Ming officially established Langboat Technology in 2021 and became an AI company incubated from scratch by Kai-Fu Lee's Innovation Works. Zhou has pointed out that Langboat Technology is dedicated to solving the challenges of human language understanding and generation, offering open-source large models based on NLP (Natural Language Processing) technology and focusing on applications and functional engines for marketing, finance, cultural creativity, and other scenarios.

Their main product is a series of capability platforms and vertical scenario applications built on the core technology of the "Mencius Large Model". They have already launched multiple technologies and products like the Mencius Large Model, the AIGC (Intelligent Creation) platform, Machine Translation platform, and Financial NLP platform which have been implemented in companies like Tonghuashun and Huaxia Fund. With technology similar to ChatGPT, Langboat Technology has launched the conversational robot MChat, which can help users complete a variety of work tasks in specific scenarios through intelligent conversation.

In March of this year, Langboat Technology completed a Pre-A+ round of financing. This round was led by Beijing Zhongguancun Science City Company, with Stodacom and Innovation Works following. In less than a year, Langboat Technology has raised several hundred million yuan in funding.

On March 14, at the trend-sharing meeting "New Opportunities from AI1.0 to AI2.0" hosted by Innovation Works, Langboat Technology officially released the "ChatGPT-like" language generation model - the Mencius MChat Controllable Large Model.

The Mencius MChat Controllable Large Model emphasizes its "controllability" - the model capability is more flexible than other similar technologies, with a focus on vertical sectors and professional tracks. It can make quick adjustments according to industry and regional demands. According to Zhou Ming, the Mencius MChat Controllable Large Model has the following features:

  • Plans to launch large models with 10B and 100B parameters;
  • Capabilities in conversation, Q&A, translation, text generation, information extraction, etc.;
  • Can integrate search results, domain data, and knowledge graphs;
  • Controllability in terms of functionality, style, human cognition, etc.

Speaking of his outlook for the future direction of the industry, Zhou Ming frankly said that ChatGPT-like technology is currently still lacking in areas like reasoning, logic, arithmetic, and factual accuracy. In the future, nine key issues related to large models are particularly worthy of attention, involving reasoning ability, factual correctness, Chinese processing capability, etc.

3. ZhiPu AI

Funding information: Series B

Established: June 11, 2019

Location: Beijing

Affiliated company: Beijing Zhipu Huazhang Technology Co., Ltd.

ZhiPu AI was founded by Professor Tang Jie of the Computer Science Department of Tsinghua University. Core members of the team have participated in the development of the "Wudao" project in collaboration with Tsinghua University and the Beijing Academy of Artificial Intelligence. In August 2022, the super-large-scale pre-training language model GLM-130B, co-developed by the Knowledge Engineering Laboratory of Tsinghua University and ZhiPu AI, officially went online. It was the only major model in Asia that was selected for the Stanford evaluation that year.

In terms of accuracy and other key indicators, GLM performed exceptionally well compared to large models from companies like OpenAI, Google Brain, and Meta. It surpassed GPT-3, Google's PaLM, and Meta's OPT large models in the MMLU, LAMBADA, BIG-bench-lite benchmark tests.

On May 16 this year, Qihoo 360 announced a strategic partnership with ZhiPu AI to co-develop the trillion-level large model "360GLM". Both sides will refer to the "Microsoft+OpenAI" cooperation model, combining large models with application scenarios.

Qihoo 360's CEO Zhou Hongyi believes that China should establish an innovative collaboration model that incorporates large tech companies and key research institutions, thereby creating a "Microsoft + OpenAI" team to lead the assault on large model technology. He stated that this cooperation with ZhiPu AI exemplifies such industry-research collaboration.

Regarding this partnership, ZhiPu AI CEO Zhang Peng said that ZhiPu AI has always stuck to its vision of making machines think like humans and realizing the concept of Model as a Service (MaaS).

Currently, the training data volume of this model is 4 trillion, with half in Chinese and half in English. It has 130 billion parameters and the training cost is between one to ten million yuan. As of May 1 this year, the model has received download and use applications from over 1000 research institutions in 69 countries.

On the basis of GLM-130B, ZhiPu AI has carried out supervised fine-tuning to get the ChatGLM model. The ChatGLM model is currently the most advanced open-source large model in China and has opened for internal testing.

Can Zhou Hongyi's dream of emulating Microsoft be realized? Can ZhiPu AI become China's next OpenAI?

4. Guang Nian Zhi Wai

Funding information: Angel+

Established: 2023

Location: Beijing

Guang Nian Zhi Wai , the AI startup co-founded by Meituan's co-founder Wang Huiwen, has attracted a lot of attention since its inception, being deemed to be "born with a silver spoon".

In February, Wang Huiwen, the founder of Mobvoi, Li Zhifei, and two partners from ZhenFund, Dai Yusen and Liu Yuan, had a meal together. The four personally witnessed the changes brought by ChatGPT, and Wang Huiwen's attitude was one of "must participate."

At the start of their conversation, Wang Huiwen took out his phone and declared his intention to join the field:

He would contribute $50 million to participate in the ChatGPT trend, hoping to join a suitable company.

Two days later, when more details about the startup were disclosed, Wang Huiwen's ideas had clearly shifted from "seeking to team up" to "I will organize."

When this lone leader announced his recruitment of "top technical talents," he assured everyone that they could focus on their talents while he would handle the trivial matters. He stressed that there was no need to worry about funding-

Apart from his personal investment of $50 million at a valuation of $200 million, the "next round of financing has already secured a subscription of $230 million from VCs"!

Afterwards, Wang Huiwen posted several job advertisements for positions like product managers, algorithm engineers, and interns.

In March, the second "big shot" who invested in Guang Nian Zhi Wai appeared-Wang Huiwen's old friend, Meituan founder Wang Xing.

Wang Xing joined the team by "putting in money," participating in the A-round investment, and serving as a director of Light Years Away. Wang Xing and Wang Huiwen were college roommates and classmates at Tsinghua University. They had worked together on several startups, creating popular brands such as Meituan.

Fighting side by side once again, Wang Xing explained his motivation in a ticket circle:

AI's large models excite me due to the enormous productivity they are about to create, but I am also concerned about their future impact on the world. Having walked the entrepreneurial path with Old Wang for nearly 20 years, I must support him as he decides to embrace this big wave.

On April 6, operations officially began at Guang Nian Zhi Wai, as updated by Wang Huiwen in his friends circle.

The reputation of the "Double Wangs" of Meituan also made it easier for "Guang Nian Zhi Wai" to raise funds in the venture capital market. After all, VCs probably also want to seize the opportunity to "recreate Meituan."

Wang Xing and Wang Huiwen, with Meituan, have weathered the thousand-group war and the hundred-group war, but it took nine years to achieve profitability for the first time. Can they last until the birth of the ChatGPT version of "Meituan"?

5.Yang ZhiLin 

In this hotly contested "arms race" of large models, there are no shortage of impressive young entrepreneurs born in the 90s, such as Yang ZhiLin, co-founder of Recursion AI, mentioned by Information.

Yang ZhiLin studied computer science at Tsinghua University as an undergraduate, under the guidance of ZhiPu AI founder Tang Jie. He graduated in 2015 with top honors in his class. Later, he pursued his Ph.D. at Carnegie Mellon University's Language Technologies Institute (LTI), the world's top-ranked institution for natural language processing (NLP) research, under Apple AI research head Ruslan Salakhutdinov and Google chief scientist William Cohen, completing his doctorate in 4 years.

During his Ph.D., Yang ZhiLin worked with Turing Award winner Yoshua Bengio to release the "HotpotQA" dataset and published XLNet and Transformer-XL, which had a significant impact on the NLP field. The work became one of the most cited papers at NeurIPS 2019 and ACL 2019, with Google Scholar citations exceeding 10,000.

In 2016, Yang ZhiLin founded Recursion AI, a company focused on using NLP, voice, multi-modal, large models, and other AI technologies to create "Sales Tech" solutions. He led multiple AI projects for ZhiPu AI and the Tsinghua research team. The "Pangu" large model was also jointly launched by the Yang ZhiLin team and Huawei Cloud.

Currently, Recursion AI has completed its B round of financing, and has achieved over 200% revenue growth for three consecutive years.

Conclusion However, the speed at which funds are burnt in the development of large AI models is also an insurmountable "natural moat" that all competitors must confront.

The training cost of OpenAI's super-large language generation model GPT-3 alone is as high as $4.6 million.

Senior AI researcher Tian Taoyuan, in an interview with the media, further explained the costs: "Training GPT3.5 once requires an expenditure of $3-4.6 million, which is just the cost of computing power and does not include the cost of talent. OpenAI has around 375 employees, with an annual wage expenditure of $200 million and an AI computing power expenditure of $500 million, requiring strong capital support."

Regardless of who ascends the throne of domestic AI startups, the new round of "bloody wind and rain" is inevitable.