Image source: Generated by Unbounded AI

Human joys and sorrows are not connected. Since the first year of artificial intelligence started in 2016, the AI ​​industry has undergone several rounds of reshuffle. With the help of ChatGPT, DeepSeek has stirred up the entire big model market like a catfish. It is also a big model startup and is regarded as a newcomer by the industry. Compared with the manufacturers of Six Little Dragons, the situation is so sunrise in the east and rain in the west.

DeepSeek shocked the industry after launching the low-cost, low-performance DeepSeek-V3, which is comparable to GPT-4o, and then released the R1 model on January 20, which topped the Apple App Store global download list six days after it was launched. The cumulative download volume of more than 110 million times in one month was launched. During this period, major cloud manufacturers quickly launched the open source version of V3 and R1, and Baidu Search, WeChat and other products are actively embracing DeepSeek.

The Kimi global reinforcement learning model k1.5 and step reasoning model Step R-mini released at the same time as DeepSeek are close to o1 in many aspects of model capabilities, but it is still submerged in DeepSeek's popular public opinion. middle.

Compared with the noise of DeepSeek, the "Six Little Dragons" also revealed one after another: Zero-100, the Dark Budget and Arbitration Case were not settled, and another executive of MIniMax resigned …

There are also frustrated VCs behind this: none of the projects supported by real money have reached the popularity of DeepSeek. At present, there have been no news of financing for four of the "Six Little Dragons" for more than half a year. In 2024, the industry said that two of the "Six Little Dragons" have fallen behind. In 2025, who will fall behind next?

Only three companies remain to continue to take root in the big model

DeepSeek's popularity It is not without signs. Since the launch of the first model, DeepSeek Coder, on November 2, 2023, more than 10 different versions of the models have been launched in more than a year. Among them, the V2 model released in May last year is comparable to the GPT-4 Turbo in performance, but the price is only 1% of that of GPT-4. Therefore, DeepSeek is called the "price butcher" and "Pinduoduo in AI", and it also set off a big model industry. The first round of price war.

On January 27, 2025, DeepSeek surpassed ChatGPT and topped the Apple APP Store free list in China and the United States, attracting global attention. What made DeepSeek so successful is its inference mockup DeepSeek-R1. According to information released by DeepSeek, R1 scored similar to the official version of o1 in many authoritative tests, and the scores in some tests exceeded the official version of o1.

In addition to the list score, open source + cost-effectiveness is the important thing that makes DeepSeek popularCombination punch. Affected by DeepSeek, former closed-source believer Baidu founder Robin Li also announced his joining the open source team. OpenAI founder Ultraman also reflected that the company has always been on the "wrong side" in its strategy in the open source field.

MiniMax, the big model "Six Little Dragons", released its first open source model on January 15. Its founder Yan Junjie also said in an interview with "Late" that "there are many experiences in starting a business for the first time. If you don't have it, if you can choose again, you should open source on the first day." Among the other five little dragons, only Zhipu was the first to open source and close source to walk on two legs. After nearly two years of hard work, the development direction of the "Six Little Dragons" has gone in a completely different way.

Zero1000 is the first basic model company to be significantly adjusted publicly. It first abolished the pre-training algorithm team and the Infra team. Some people joined Alibaba in the form of job-changing, and then announced that they would go with Alibaba Cloud and Suzhou. The High-tech Zone jointly established an industrial large-scale model joint laboratory and an industrial large-scale model base.

In terms of personnel, Huang Wenhao, the person in charge of model training, Lan Yuchuan, the person in charge of the open platform of the big model API, and Cao Dapeng, the person in charge of the productivity product, all resigned one after another. The zero-one things that try to stay on the card table cannot cover up the decline in this round of mock-up competition.

Baichuan Intelligent has made it clear that it will take the medical track in 2024 and recently launched its first "AI pediatrician". The commercialization of To B seems to be not very smooth. Hong Tao, its co-founder and head of commercialization, had already resigned a year ago. According to an employee of Baichuan, it was indeed not as expected, "Now with DeepSeek, the pressure this year has only increased and not decreased."

The head of To B commercialization also left Wei Wei of MiniMax, who had previously Wei Wei. Wei said in an interview that many B-side customers will not pay this money easily to support the revenue of large-scale model companies. They can only help customers align the output effects in actual scenarios based on R&D capabilities and algorithm capabilities, which also confirms the large-scale model. Commercialization is not easy.

In this way, the only ones who are still focusing on big model technology innovation and pursuing AGI are the dark side of the moon, the wisdom spectrum, and the leap of the stars. Influenced by DeepSeek, Step Yuexingchen has also joined the open source camp. However, unlike DeepSeek's focus on text models, Step Yuexingchen's latest open source are two multimodal models - Step-Video-T2V and Step-Audio .

In the early morning of February 23, the Dark Side of the Moon released the latest paper "Muon is Scalable for LLM Training", and opened the source MoE model Moonlight, with only 3B activation parameters required for the model. Many industry insiders believe that this is the "Intercepting the Husband and Open Source Week", because DeepSeek earlier announced that it would release open source projects for five consecutive days.

For the dark side of the moon, the one that burns the eyebrows may be its massive Kimi product.

Buying money and investing in traffic is difficult to become the top brother

Like the big model "Six Xiaolongs", DeepSeek also has a C-end product of the same name, which did not attract too much attention in the market in the first week after its launch. According to data disclosed to the media by QuestMobile, from January 13 to January 19, 2025, the weekly download volume of DeepSeek App was only 285,000, far less than Doubao (4.52 million) and Kimi (1.557 million).

After the release of R1 on January 20, 2025, DeepSeek downloads began to grow steeply. Sensor Tower research shows that DeepSeek downloads exceeded 16 million times within 18 days of the press conference, almost the first time OpenAI's ChatGPT is released. twice as many as 9 million times.

The surge in visits once caused DeepSeek to crash. Even so, the growth momentum is still very strong, with monthly downloads exceeding 110 million. No one of the DeepSeek's brilliance can be ignored. On February 13, when CEO Liang Rubo talked about DeepSeek, he reflected on the lack of follow-up speed and pursued intelligent launch this year.

Tencent's WeChat grayscale test is connected to DeepSeek's AI search. After the usage volume exceeds expectations, it calls the AI ​​application Yuanbao to support WeChat search. On February 22, Tencent Yuanbao surpassed Bytes' bean bag and rose to the second place in the Chinese Apple free APP download rankings, and DeepSeek continued to top the list.

The "Brother One and Two" changed hands in just one month, forcing money to exchange for growth bean bags and Kimi no longer have the advantages. The difference between the two is that the former is a noble born with a "golden key" and the latter is a "new entrepreneurial entrepreneurship". Previously, media calculated that Kimi's daily investment amount was close to 200,000 yuan on the iPhone channel alone, while Doubao was 2.48 million yuan.

Under the influence of DeepSeek, the Dark Side of the Moon has been reported to have significantly reduced its product placement budget, including suspending the placement of multiple Android channels and cooperation with third-party advertising platforms. According to insiders to AI Light Years, the promotion has indeed made corresponding adjustments, "There are natural new additions, but they cannot be compared with the growth trend of DeepSeek."

kimi's current worries are more than these: "Undercover Waves" exclusively learned that the Kimi arbitration case, which had been shelved for a long time, did not complete the settlement as expected, but entered the next process of the arbitration case. According to insiders, the two parties to the Kimi arbitration case, the old shareholder of Cycle Intelligence and Yang Zhilin, have completed the payment at HKIAC (Hong Kong International Arbitration Center) in the end of January and late February respectively, and the party has been completed. The more critical protagonist behind the entire incident, Zhang Yutong, may be filed separately.

MiniMax also has high expectations for To C products because its celebrity product Talkie202In the first half of 4 years, it became the fourth largest AI application in the United States, making it tasted good. But the good times didn't last long. Talkie quietly disappeared in the Apple app store in the US market in mid-December, while the Android platform was unaffected.

Skills, Zero One Wanwu, Zhipu AI and Baichuan Intelligent also have their own AI application products, but according to the AI ​​product list, it shows that in January 2025, there will be the top 20 AI applications in the top 20 monthly active users. None of them are related to these four manufacturers. Previously, Baichuan Intelligent's employees told AI Light Years, "It is not surprising that Bai Xiaoying's user retention and growth are poor. We basically do not advertise and let other families spend money to complete user education first."

< p> Currently, DeepSeek, Tencent Yuanbao and ByteDance are the top three in Apple's free APP download rankings. If the big model "Six Little Dragons" wants to be on the list, the competition will only be more intense. Zhou Hongyi, currently ranked seventh, is personally taking the stage to "bring goods".

Another rival that cannot be ignored is Alibaba. After the AI ​​application Tongyi was incorporated into the Alibaba Intelligent Information Business Group, Alibaba's AI To C business recently launched large-scale recruitment, with hundreds of positions, and concentrated In product and technology research and development positions related to AI big model. There are wolves in front and tigers behind, which is a true portrayal of the current situation of the big model "Six Little Dragons".

When the technology story is no longer romantic, commercialization is less than expected, and the growth of monthly active users of the product is not proportional to investment, the big model "Six Little Dragons" is ideal and full, and realistic.

The threshold for the next round of financing is raised

The pre-training of large models is a recognized fact, Kai-Fu Lee once revealed, The cost of pre-training is about three to four million US dollars. Even Yi-Lightning, which is cheaper, used 2,000 GPUs during training, which took a month and a half and cost more than three million US dollars.

Even DeepSeek, which claims to be low-cost, has an incalculable investment in the early stage. According to the third-party agency SemiAnalysis, DeepSeek actually has a huge computing power reserve: a total of 60,000 Nvidia GPU cards have been built, including 10,000 A1 million, 10,000 H1 million, 10,000 "special version" H800 and 30,000 "special version" H20.

"The training cost of a general-purpose big model is estimated to be about 1 billion US dollars. This is only the computing power part, and the other two very expensive parts are not counted. One is data and the other is Human resources cost, talents in the global big model field are now very scarce. "Dr. Du Feng, founding partner of Jiangmen Venture Capital and former head of Microsoft Venture Capital Greater China, once told the author.

Because such high investment is required, a saying has been popular in the industry for a long time: the ticket to investing in big model companies is US$100 million. Another signal behind this sentence is that if a big model startup company cannot get financing, it will be difficult for it to survive.Go.

After the 100-model war in 2023, financing news will be released almost every other month, but with the popularity of AI bubble theory, starting from September 2024, there have been no hundreds of millions of dollars for a long time. Hot money flows to the big model "Six Little Dragons". Until before the Spring Festival in 2025, Zhipu and Jieyuexingchen announced that they had obtained "winter money". The former announced that they had completed a new round of financing of RMB 3 billion, while the latter completed a round of financing of hundreds of millions of dollars.

The other four of the "Six Little Dragons" have been more than half a year since the last financing trend was released: MiniMax officially announced its completion of a US$600 million Series B financing in March last year, and Baichuan Intelligent received 5 billion yuan in A in July last year The round of financing, Zero-100 million yuan in August last year, and the monthly dark side completed a $300 million yuan in financing last August.

During the Spring Festival, DeepSeek became popular all over the world, and public opinion praised DeepSeek and its founder Liang Wenfeng without hesitation. In the venture capital circle, a lot of news has been circulating recently about whether DeepSeek will start financing and how much valuation it is.

There was previously reported that Alibaba would be valued at US$10 billion, with an investment of US$1 billion to account for 10% of the shares. In response, Alibaba Vice President Yan Qiao quickly denied the rumors through his circle of friends, saying that "the information about Alibaba's investment in DeepSeek is fake news." Later, foreign media said that "DeepSeek is considering raising external funds for the first time." DeepSeek related people refuted the rumors, and the financing news was all rumor.

"Many investors directly or rely on their relationship to Liang Wenfeng. I predict that the valuation should be far beyond the current 'big model Six Xiaolong'." An investor in CICC said, " DeepSeek has become the benchmark target. The threshold for Liu Xiaolong to obtain new financing in the primary market is obviously higher. "

In fact, since the wave of big model entrepreneurship has been launched, the industry generally does not believe it. The "Six Little Dragons" can eventually survive as an independent "big model company". Several founders of the "Six Little Dragons" have also expressed similar views in public, such as MiniMax founder Yan Junjie, who believes that there will only be five large-scale companies left in the world in the future.

"China must have its own ChatGPT. This is like search engines, and we have our own compliance requirements. But the Chinese version of ChatGPT will only be generated in five companies: BAT+Byte+Huawei. "Cheng Hao, founder of Thunder and Yuanwang Capital, once told the author.

With the continued popularity, the "Six Little Dragons" that are already moving towards differentiation will accelerate the reshuffle.