Under the Glory of DeepSeek: The Forlorn "Six Little Dragons"
TechFlow Selected TechFlow Selected
Under the Glory of DeepSeek: The Forlorn "Six Little Dragons"
Under the continued surge of DeepSeek, the already-diverging "Six Little Dragons" will accelerate their reshuffling.
Navigating Web3 tides with focused insightsAuthor: Wu Qianyu
Human joys and sorrows do not resonate with one another. Since the beginning of the AI era in 2016, the AI industry has undergone several rounds of reshuffling. Riding on the wave of ChatGPT, DeepSeek has stirred up the entire large model market like a catfish. Compared to other large-model startups hailed as the new "Six Dragons," fellow players have seen vastly different fates—some basking in sunshine while others drenched in rain.
After shocking the industry with its low-cost DeepSeek-V3 model, comparable in performance to GPT-4o, DeepSeek launched its R1 model on January 20. Six days after launch, it topped the global download chart on Apple's App Store, reaching over 110 million cumulative downloads within a month. During this period, major cloud providers quickly rolled out open-source versions of V3 and R1, while products such as Baidu Search and WeChat actively integrated DeepSeek.
In contrast, Kimi’s globally scaled reinforcement learning model k1.5 and Jieyue’s reasoning model Step R-mini, despite achieving capabilities close to o1 across multiple dimensions, were still drowned out by the overwhelming media attention surrounding DeepSeek.
Compared to DeepSeek’s hype, the "Six Dragons" have been hit with consecutive setbacks: Zhipu split further, Moonshot faced unresolved budget disputes and arbitration cases, and MiniMax saw yet another executive departure…
Behind them are also disillusioned VCs: none of the projects they backed with real capital have reached the level of buzz achieved by DeepSeek. Currently, four out of the “Six Dragons” have gone more than half a year without releasing any new funding news. In 2024, two were already considered lagging behind; in 2025, who will be the next?
Only Three Remain Focused on Large Models
DeepSeek’s breakout was not unanticipated. Since launching its first model, DeepSeek Coder, on November 2, 2023, it has released over ten different versions in just over a year. Its V2 model, launched last May, matched GPT-4 Turbo in performance but cost only 1% of GPT-4, earning DeepSeek nicknames like "price butcher" and "Pinduoduo of AI," while triggering the first price war in the large model industry.
On January 27, 2025, DeepSeek surpassed ChatGPT to top the free app charts in both China and the U.S. on Apple’s App Store, drawing global attention. The driving force behind this achievement is its reasoning model, DeepSeek-R1. According to publicly available data, R1 scored nearly on par with the official version of o1 in multiple authoritative benchmarks, even exceeding it in some tests.
Beyond benchmark scores, the combination of open-sourcing and high cost-effectiveness has been key to DeepSeek’s explosive popularity. Under pressure from DeepSeek, even Li Yanhong, founder of Baidu—a long-time advocate of closed-source models—announced his move toward open source. Sam Altman, OpenAI’s CEO, reflected that the company had been on the “wrong side” of open-source strategy.
MiniMax, one of the “Six Dragons,” released its first open-source model on January 15. Founder Yan Junjie told Wanliang that if he could start over, “I would open-source from day one—we lacked experience during our first venture.” Among the other five, only Zhipu adopted a dual-track approach of both open and closed source from early on. After nearly two years of struggle, the strategic directions of the “Six Dragons” have diverged significantly.
Zhipu was the first foundational large model company to publicly make major adjustments—first cutting its pre-training algorithm team and Infra team, with some members joining Alibaba via job transfers, then announcing joint ventures with Alibaba Cloud and Suzhou High-Tech Zone to establish industrial large model labs and bases.
On the personnel front, Huang Wenhao (in charge of model training), Lan Yuchuan (responsible for the large model API platform), and Cao Dapeng (productivity product lead) all left consecutively. Despite attempts to stay in the game, Zhipu cannot hide its weakening position in this round of large model competition.
Baichuan Intelligence clarified its focus on the healthcare sector in 2024 and recently launched its first "AI pediatrician." However, its B2B commercialization efforts seem less successful—its co-founder and head of commercialization, Hong Tao, left before the Lunar New Year. An employee at Baichuan confirmed expectations were unmet: “Now that DeepSeek exists, the pressure this year only grows.”
MiniMax also lost Wei Wei, its B2B commercialization lead. In a prior interview, Wei noted that many enterprise clients are unwilling to pay enough to sustain large model companies’ revenue, so companies must rely on R&D and algorithmic capabilities to align outputs with client needs in practical scenarios—confirming how difficult large model monetization truly is.
Thus, only Moonshot, Zhipu, and StepFun remain committed to large model technological innovation and AGI pursuit. Influenced by DeepSeek, StepFun joined the open-source movement—but unlike DeepSeek’s focus on text models, StepFun’s latest open-sourced offerings are two multimodal models: Step-Video-T2V and Step-Audio.
In the early hours of February 23, Moonshot released a new paper titled "Muon is Scalable for LLM Training" and open-sourced its MoE model Moonlight, requiring only 3B activated parameters. Many industry insiders believe this was an attempt to "scoop the open-source week," as DeepSeek had previously announced plans to release open-source projects daily for five consecutive days.
For Moonshot, however, the most pressing concern may be its heavily funded Kimi product.
Burning Cash on User Acquisition Falls Short of Top Spot
Like the other "Six Dragons," DeepSeek also has a consumer-facing app bearing its name, which initially drew little market attention during its first week post-launch. According to data disclosed by QuestMobile to media, from January 13 to January 19, 2025, the weekly download volume of the DeepSeek app was only 285,000—far below Doubao (4.52 million) and Kimi (1.557 million).
However, after the release of R1 on January 20, 2025, DeepSeek’s download numbers began a steep climb. Sensor Tower research shows that within 18 days of the launch, DeepSeek accumulated over 16 million downloads—nearly double the 9 million downloads OpenAI’s ChatGPT achieved upon its initial release.
The surge in traffic briefly caused DeepSeek’s service to crash, yet growth remained strong, surpassing 110 million monthly downloads. DeepSeek’s brilliance can no longer be ignored. At ByteDance’s all-hands meeting on February 13, CEO Liang Rubo reflected that their response speed was insufficient and emphasized pursuing intelligent deployment this year.
Tencent’s WeChat began gray-testing integration with DeepSeek’s AI search. After usage exceeded expectations, Tencent deployed its own AI app Yuanbao to support WeChat search. On February 22, Tencent’s Yuanbao surpassed ByteDance’s Doubao to rank second on the Chinese App Store’s free app download chart, while DeepSeek remained at number one.
The rapid shift in rankings between the "top two brothers" within just a month has eroded the once-dominant positions of Doubao and Kimi, both of which relied on heavy spending for growth. The difference lies in their origins: one born a noble with a "golden key," the other a "startup newcomer." Media reports estimated that Kimi spent nearly 200,000 yuan daily on iPhone channel ads alone, compared to Doubao’s 2.48 million.
Under DeepSeek’s influence, Moonshot was recently reported to have drastically cut its product promotion budget, including suspending campaigns on multiple Android channels and third-party advertising platforms. An insider told AI Guangnian: “There’s organic growth, but nothing compares to DeepSeek’s momentum.”
Kimi’s troubles go beyond this: Waves exclusively learned that the long-pending Kimi arbitration case did not reach a settlement as expected, but instead advanced into the next phase. According to sources, both parties—the former shareholders of Xunhuan Intelligence and Yang Zhiyun et al.—completed payment at HKIAC (Hong Kong International Arbitration Centre) by the end of January and late February respectively, and the tribunal has now been formed. Zhang Yutong, a central figure behind the incident, may face a separate lawsuit.
MiniMax also places high hopes on its consumer product. Its star app Talkie ranked fourth among downloaded AI apps in the U.S. in the first half of 2024, giving the company a taste of success. But the good times didn’t last. In mid-December, Talkie quietly disappeared from Apple’s U.S. App Store, though it remained accessible on Android.
StepFun, Zhipu, Baichuan, and Zero One All Things also have their own AI applications. However, according to the AI Product Ranking, none of these four companies made it into the top 20 most active AI apps in January 2025. A former Baichuan employee told AI Guangnian: “It’s no surprise that Baixiaoying has poor user retention and growth—we don’t run ads and are letting others spend money to educate users first.”
Currently, DeepSeek, Tencent Yuanbao, and ByteDance Doubao occupy the top three spots on Apple’s free app download chart. For the “Six Dragons,” breaking onto the list means fiercer competition. Nanos Research, currently ranked seventh, now sees Zhou Hongyi personally stepping in to promote it.
Another formidable competitor is Alibaba. After merging Tongyi into Alibaba’s Smart Information Group, Alibaba recently launched large-scale hiring for its AI-to-consumer business, posting hundreds of roles focused on AI large model-related product and technical development. With wolves ahead and tigers behind, this perfectly captures the current predicament of the “Six Dragons.”
When technical narratives lose their charm, commercialization falls short, and product MAU growth fails to match investment, the ideals of the “Six Dragons” remain lofty, but reality is harsh.
Next Round of Funding Bar Raised
It's widely acknowledged that pre-training large models is extremely costly. Li Kaifu once revealed that a single pre-training session costs around $3–4 million. Even Yi-Lightning, known for lower costs, used 2,000 GPUs, took a month and a half, and cost over $3 million.
Even DeepSeek, despite branding itself as low-cost, required massive upfront investment. Third-party firm SemiAnalysis estimates DeepSeek possesses enormous computing power: approximately 60,000 Nvidia GPU cards, including 10,000 A100s, 10,000 H100s, 10,000 “China-specific” H800s, and 30,000 “China-specific” H20s.
"We estimate the cost of training a general-purpose large model at around $1 billion—this covers only compute, excluding two other major expenses: data and labor. Talent in the global large model field is extremely scarce," Dr. Du Feng, founding partner at Jiayuan Capital and former Microsoft Ventures China head, told the author.
Due to such high costs, a common saying circulated in the industry for a long time: the entry ticket for investing in large model startups is $100 million. Another signal behind this statement is clear: without funding, a large model startup cannot survive.
After the “hundred-model battle” erupted in 2023, funding announcements emerged almost monthly. But as skepticism about the AI bubble grew louder, from September 2024 onward, no significant investments flowed into the “Six Dragons” for an extended period. It wasn't until shortly before the 2025 Spring Festival that Zhipu and StepFun announced securing their "winter survival funds"—Zhipu completed a new round of financing worth 3 billion RMB, while StepFun secured hundreds of millions in Series B funding.
The other four among the “Six Dragons” have not announced funding updates for over six months: MiniMax announced its $600 million Series B in March last year, Baichuan secured 5 billion RMB in Series A in July, Zero One All Things raised hundreds of millions in a new round in August, and Moonshot completed a $300 million funding round in August.
During the Spring Festival, DeepSeek went viral globally, with public praise lavished on both DeepSeek and its founder Liang Wenfeng. In venture circles, rumors have been circulating about whether DeepSeek will raise funds and at what valuation.
Rumors suggested Alibaba would invest $1 billion for a 10% stake at a $10 billion valuation. Alibaba VP Yan Qiao quickly debunked this via social media, stating, “The rumor about Alibaba investing in DeepSeek is false.” Later, foreign media claimed “DeepSeek is considering raising external capital for the first time,” which DeepSeek insiders also denied, calling all funding reports baseless.
"Many investors are directly or indirectly trying to meet Liang Wenfeng. I predict its valuation will far exceed that of the current 'Six Dragons,'" said an investor from CICC Capital. "DeepSeek has become the benchmark—raising new funds in the primary market now has a much higher bar."
In fact, since the rise of large model startups, few in the industry believed all six could survive independently as standalone "large model companies." Founders among the “Six Dragons” have expressed similar views publicly—for instance, MiniMax’s Yan Junjie believes only five large model companies will remain globally.
"China will definitely have its own ChatGPT—just like search engines, we have our own compliance requirements. But China’s version of ChatGPT will only emerge from five companies: BAT, ByteDance, and Huawei," Cheng Hao, founder of Xunlei and Far Reach Capital, told the author.
Amid sustained momentum, the already-diverging “Six Dragons” will accelerate their shakeout.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
