Author: Wu Qianyü
The joys and sorrows of humans are not universally shared. Since the launch of the AI element year in 2016, the AI industry has already experienced several rounds of reshuffling. Riding on the momentum of ChatGPT, DeepSeek has stirred up the entire large model market like a catfish. The manufacturers that are also large model startup companies and are regarded as the new "Six Small Dragons" by the industry have a fate that can be described as "the sun rises in the east and the rain falls in the west" compared to DeepSeek.
After DeepSeek released the low-cost DeepSeek-V3 model, which performs on par with GPT-4o, before the Lunar New Year, it then released the R1 model on January 20th. The R1 model topped the global download chart of the Apple App Store within six days of its launch, and accumulated over 110 million downloads within a month. During this period, major cloud vendors quickly launched open-source versions of V3 and R1, and products like Baidu Search and WeChat actively embraced DeepSeek.
Meanwhile, the global reinforcement learning model k1.5 and the step reasoning model Step R-mini released by Kimi around the same time as DeepSeek, while being close to o1 in many aspects of model capability, were still overshadowed by the heated discussions around DeepSeek.
Compared to DeepSeek's clamor, the "Six Small Dragons" have also been continuously releasing news: Zero One Wanwu further divides, the Lunar Dark Side's budget and arbitration case remain unresolved, and another senior executive has left MIniMax...
Behind this are also the disappointed VCs: not a single project they have supported with real money has reached the level of heat that DeepSeek has. Currently, more than half of the "Six Small Dragons" have not released any financing news for over half a year. The industry predicts that two of the "Six Small Dragons" will fall behind in 2024, and the question is, who will be the next to fall behind in 2025?
Only three companies continue to root in large models
DeepSeek's blazing success was not without signs. Since launching its first model DeepSeek Coder on November 2, 2023, it has released over 10 different model versions in just over a year. The V2 model released last May performed on par with GPT-4 Turbo, but was priced at only 1% of GPT-4, earning DeepSeek the titles of "price butcher" and "AI Pinduoduo", and also sparked the first round of price wars in the large model industry.
On January 27, 2025, DeepSeek surpassed ChatGPT and topped the free app download charts in the Chinese and US Apple App Stores, drawing global attention. The key to DeepSeek's success is its reasoning large model DeepSeek-R1. According to DeepSeek's information, R1 scored close to the official version of o1 in multiple authoritative tests, and even surpassed the official version of o1 in some tests.
In addition to the chart rankings, the combination of open-source and cost-effectiveness is the important one-two punch that has ignited the tremendous heat around DeepSeek. Affected by DeepSeek, Baidu founder Li Yanhong, who was once a closed-source believer, also announced to join the open-source camp, and OpenAI founder Sam Altman also reflected that the company has been "on the wrong side" in its open-source strategy.
Among the large model "Six Small Dragons", MiniMax released its first open-source model on January 15th, and its founder Yan Junjie said in an interview with "Wanpoint" that "there were many experiences lacking in the first startup, and if I could choose again, I should have open-sourced from the first day." Of the other five small dragons, only Zhipu was the earliest to walk on two legs of open-source and closed-source. After nearly two years of groping and struggling, the development direction of the "Six Small Dragons" has become completely divergent.
Zero One Wanwu is the first company to publicly make major adjustments to its basic large model. It first laid off the pre-training algorithm team and the Infra team, with some personnel joining Alibaba in the form of job-hopping, and then announced the establishment of an industrial large model joint laboratory with Alibaba Cloud and the Suzhou High-tech Zone, as well as an industrial large model base.
In terms of personnel, the model training leader Huang Wenhao, the person in charge of the large model API open platform Lan Yuechuan, and the head of productivity products Cao Dapeng have all resigned in succession. Zero One Wanwu, which is trying to stay on the game board, cannot conceal its decline in this round of large model competition.
Baichuang Intelligence also clearly defined its medical track in 2024 and recently launched its first "AI pediatrician". However, Baichuang's commercialization in the B-end does not seem to be going too smoothly, and its co-founder and commercial director Hung Tao had already left before the Lunar New Year. According to an employee of Baichuang, the results are not as expected, "Now that we have DeepSeek, the pressure this year will only increase, not decrease."
The person in charge of B-end commercialization who has also left is Wei Wei of MiniMax. In a previous interview, Wei Wei said that many B-end customers would not easily pay the money to support the revenue of large model companies, and could only help customers align the output effect in actual scenarios based on R&D capabilities and algorithm capabilities, which also proves that the commercialization of large models is not an easy task.
It seems that the only ones still focused on large model technology innovation and pursuing AGI are Lunar Dark Side, Zhipu, and Juecestar. Affected by DeepSeek, Juecestar has also joined the open-source camp, but unlike DeepSeek's focus on text models, Juecestar's latest open-source models are two multimodal models - Step-Video-T2V and Step-Audio.
In the early hours of February 23rd, Lunar Dark Side released the latest paper "Muon is Scalable for LLM Training" and open-sourced the MoE model Moonlight, with an activation parameter of only 3B. Many industry insiders believe this is a "preemptive open-source" move, as DeepSeek had previously announced that it would release open-source projects for five consecutive days.
For Lunar Dark Side, the burning issue may be its large-scale investment in the Kimi product.
Burning money for traffic fails to become the top dog
Like the large model "Six Small Dragons", DeepSeek also has a C-end product of the same name, which did not attract much attention in the market in the first week after its launch. According to data disclosed by QuestMobile to the media, from January 13 to January 19, 2025, the weekly download volume of the DeepSeek App was only 285,000, far behind Douban (4.52 million) and Kimi (1.557 million).
After the release of R1 on January 20, 2025, the download volume of DeepSeek began to grow steeply. Sensor Tower research shows that within 18 days of the release, DeepSeek's download volume exceeded 16 million, nearly twice the 9 million downloads of OpenAI's ChatGPT at its initial release.
The surge in traffic once caused DeepSeek to crash, but the growth momentum is still very strong, with a monthly download volume exceeding 110 million. DeepSeek's brilliance can no longer be ignored by anyone, and on February 13, Bytedance CEO Liang Rubo reflected on the company's speed of catching up in an internal all-hands meeting, saying that this year they need to pursue intelligent online.
WeChat of Tencent has grayed in the integration of DeepSeek's AI search, and after the usage exceeded expectations, it also called in the AI application Yuanbao to support WeChat search. On February 22, Tencent's Yuanbao surpassed Bytedance's Douban and rose to the second place in the free app download ranking on the Chinese Apple App Store, with DeepSeek continuing to top the list.
The "top dog number one and number two" have changed hands in just one month, forcing the advantages of Douban and Kimi, which burn money for growth, to no longer exist. The difference is that the former is a "noble" born with a "golden key", while the latter is a "new rich" startup.
Under the influence of DeepSeek, Lunar Dark Side has recently been exposed to have significantly reduced its product promotion budget, including suspending placements on multiple Android channels and cooperation with third-party advertising platforms. According to an internal source revealed to "AI Lightyear", the promotion has indeed been adjusted accordingly, "There is natural growth, but it is impossible to compare with DeepSeek's surge."
Kimi's current troubles are not just these: "Undercurrent Waves" has learned that the long-pending Kimi arbitration case has not been resolved as expected, but has entered the next stage of the arbitration process. According to an informed source, the two parties in the Kimi arbitration case, the old shareholders of Circular Intelligence and Yang Zhilin, have respectively completed the payment of fees at the HKIAC (Hong Kong International Arbitration Center) at the end of January and the end of February, and the formation of the tribunal has also been completed. And the key figure behind the whole event, Zhang Yutong, may be sued separately.
MiniMax, which also has high hopes for its C-end product, is because its flagship product Talkie became the fourth most downloaded AI application in the US in the first half of 2024. But the good times did not last long, as Talkie quietly disappeared from the Apple App Store in the US market in mid-December, while the Android platform was not affected.
The stars and moons, the zero and the myriad things, the AI map and the Baichuan Intelligence all have their own AI application products, but according to the AI product list, none of the top 20 AI applications in January 2025 are related to these four manufacturers. Previously, an employee of Baichuan Intelligence told "AI Light Year" that "the user retention and growth of Bai Xiao is not surprising, we basically don't do advertising, let the others spend money to complete user education first."
Currently, DeepSeek, Tencent Yuanbao, and Byte Dou Bao occupy the top three positions on the Apple free APP download ranking. The "six small dragons" of the large model want to make the list, and the competition will only become more intense. The seventh-ranked Nano Search, led by Zhou Hongyi, is personally "bringing goods".
Another opponent that cannot be ignored is Alibaba. After the AI application Tongyi was merged into Alibaba's Smart Information Business Group, Alibaba's AI To C business has recently launched a large-scale recruitment, with hundreds of positions focused on AI large model-related products and technology R&D. With wolves in front and tigers behind, this is the true portrayal of the current situation of the "six small dragons" of the large model.
When the technology story is no longer romantic, the commercialization is not as expected, and the monthly active users and investment of the product are not proportional, the "six small dragons" of the large model have ideals that are full of dreams, but the reality is bone-dry.
The threshold for the next round of financing has been raised
The fact that large model pre-training burns money is a recognized fact. Li Kaifu once revealed that the cost of a single pre-training is about $3-4 million. Even the lower-cost Yi-Lightning used 2,000 GPUs during training, taking one and a half months and costing more than $3 million.
Even the low-cost DeepSeek, the upfront investment is difficult to estimate. The third-party agency SemiAnalysis estimates that DeepSeek actually has a huge computing power reserve: a total of 60,000 NVIDIA GPU cards, including 10,000 A100s, 10,000 H100s, 10,000 "special edition" H800s, and 30,000 "special edition" H20s.
"The training cost of general large models, we estimate is around $1 billion, which is just the computing power part, not counting the other two very expensive parts, one is data, and the other is labor costs. The talent in the global large model field is very scarce," Dr. Du Feng, founding partner of Zhi Men Venture Capital and former head of Microsoft Venture Greater China, told the author.
Due to the need for such high investment, a saying has become popular in the industry for a long time: the entry ticket for investing in large model companies is $100 million. The other signal behind this sentence is that a large model startup company cannot survive without financing.
After the "hundred model war" broke out in 2023, financing news was released almost every month, but with the AI bubble theory becoming increasingly prominent, from September 2024, there was no hot money of hundreds of millions flowing to the "six small dragons" of the large model for a long time. Until just before the 2025 Spring Festival, Zhipu and Jiaoyue Xingchen announced that they had obtained the "winter money", with the former announcing the completion of a new round of 3 billion yuan financing and the latter completing a B round of hundreds of millions of dollars.
The other 4 companies in the "six small dragons" have been more than half a year since the last financing dynamics were announced: MiniMax officially announced the completion of a $600 million B round of financing in March last year, Baichuan Intelligence received a 5 billion yuan A round of financing in July last year, Zero and All Things completed a new round of hundreds of millions of dollars in financing in August last year, and Moonlight completed a $300 million financing in August last year.
During the Spring Festival, DeepSeek became a global sensation, and the media did not spare praise for DeepSeek and its founder Liang Wenfeng. In the investment circle, there have been a lot of rumors circulating recently about whether DeepSeek will start financing and what the valuation will be.
Previously, there was news that Alibaba would invest $1 billion for a 10% stake at a $10 billion valuation. Alibaba's vice president Yan Qiao quickly denied this through his circle of friends, saying that "the information circulating about Alibaba investing in DeepSeek is false." Later, foreign media reported that "DeepSeek is considering raising external funds for the first time," and DeepSeek-related personnel denied the financing news as rumors.
"Many investors have directly or through connections tried to contact Liang Wenfeng, and I predict the valuation should be far higher than the current 'six small dragons' of the large model," said an investor at CICC. "DeepSeek has become the benchmark, and the six small dragons will have a higher threshold to obtain new financing in the primary market."
In fact, since the large model entrepreneurial wave started, the industry generally does not believe that the "six small dragons" will all be able to survive as independent "large model companies" in the end. Some of the founders of the "six small dragons" have also expressed similar views in public, such as MiniMax founder Yan Junjie believing that there will only be 5 large model companies left in the future globally.
"China will definitely have its own ChatGPT. Just like search engines, we have our own compliance requirements. But the Chinese version of ChatGPT will only be produced by 5 companies: BAT + Byte + Huawei," Xunlei founder and Yuanwang Capital Cheng Hao told the author.
Under the sustained fire, the "six small dragons" that are already heading towards differentiation will accelerate the reshuffle.
