Liang Wenfeng retained 97% of the employees.

“There is indeed contact regarding DeepSeek's financing,” a friend from an FA institution told us.

In the past half - month, the news that DeepSeek has finally started its financing has spread. When the outside world attributes this “abnormality,” almost inevitably, they mention that Liang Wenfeng is going to give a definite valuation to the core internal employees.

After all, in the past year, the competition in the large - model field has become increasingly intense, and there has been a lot of talent flow at DeepSeek. For example, Luo Fuli, Wang Bingxuan, and Guo Daya have successively jumped to Xiaomi, Tencent, and ByteDance.

Beyond the hustle and bustle, a set of data reflects a more peaceful situation — DeepSeek V4 was finally released in late April. A long list of author acknowledgments in the technical report shows that 10 out of about 270 people in the research and engineering team left during the R & D period. Correspondingly, the turnover rate of R & D personnel is less than 4%.

Actually, Liang Wenfeng has retained the vast majority of people.

Unveiling the Job - hopping Storm at DeepSeek: 10 Employees Left

Since 2023, a strong sense of impetus has hit.

After the sweep of ChatGPT, star companies such as Moon's Dark Side, Stepping Stars, and MiniMax were established either in this year or around this time. Large - model products of big companies like Doubao, Tongyi Qianwen, and Wenxin Yiyan have emerged intensively.

Right on time, Liang Wenfeng also launched DeepSeek in Beijing and Hangzhou this year.

In a rare public sharing at this time, he talked about his view on talent: Most of DeepSeek's developers are fresh graduates or those with short - term AI work experience. If pursuing short - term goals, it is of course right to recruit experienced people. But in the long run, basic skills, creativity, and enthusiasm are more important.

Indeed. In early 2025, when DeepSeek R1 showed its strength, people began to really pay attention to this team of about 150 people. Many of them are young people who have just graduated or are still studying at top domestic universities, with a high proportion from Peking University and Tsinghua University.

Almost inevitably, in the following year, the topic of talent flow began to hit DeepSeek.

Since 2025, it has been reported that core backbones such as Luo Fuli, Wang Bingxuan, Wei Haoran, and Ruan Chong have left DeepSeek. Many of them have jumped to other places and become core business leaders. There is a popular joke in the industry: “When members inside DeepSeek find that people of similar levels can get so much after job - hopping, why can't I?”

Until early 2026, with Guo Daya's job - hopping to ByteDance's seed team, the discussion about talent loss at DeepSeek reached a peak. Coupled with the long - awaited release of DeepSeek V4 at that time, it was inevitable to raise some concerns about a possible talent shortage.

But the reality is not so discouraging. Now, DeepSeek V4 has finally made its debut. A list of author acknowledgments was disclosed in the simultaneously released technical report. Looking closely, there are about 270 people in the Research & Engineering team, which is considered the most core R & D team of an AI company. There are also 48 people in the Business & Compliance team.

During the R & D period of DeepSeek V4, only 10 members of the research and engineering team left.

That is to say, 10 out of 270 people in the R & D team chose to leave, and the turnover rate of the core department is only less than 4% — which is quite low. A set of data shows that OpenAI lost more than 25% of its key research talents in the past two years. Most of them jumped to competitors like Meta or started their own businesses.

Opening the Financing Door for the First Time to Stabilize Morale

Currently, the venture capital circle is especially looking forward to: Who can participate in DeepSeek's first - round financing?

Since April, it was first reported that DeepSeek is launching its first - round external financing at a valuation of over $10 billion. Just one week later, it was reported that DeepSeek is in talks with Tencent and Alibaba about investment. Later, it was rumored in the industry that DeepSeek's pre - investment valuation is 300 billion RMB.

As of now, DeepSeek has not made any response to the financing news.

An FA told us that recently they have been in contact with investment institutions about the cooperation method for DeepSeek's financing. There are very few financial investment institutions in this round of financing. Another point has also been confirmed: Tencent has communication with DeepSeek in daily business, but there is no substantial negotiation on financing.

Everything is still shrouded in mystery.

On April 27, DeepSeek's registered capital increased from 10 million RMB to 15 million RMB. Among them, the registered capital subscribed by Liang Wenfeng increased from 100,000 RMB to 5.1 million RMB, and his direct shareholding ratio increased from 1% to 34%. Synchronously, the shareholding ratio of Ningbo Cheng'en Enterprise Management Consulting Partnership, controlled by Liang Wenfeng, decreased from 99% to 66%. After this change, Liang Wenfeng holds about 84.29% of DeepSeek's equity indirectly and directly.

It is worth noting that previously, Liang Wenfeng held the vast majority of DeepSeek's equity through Ningbo Cheng'en, with very little direct shareholding. After this change, Liang Wenfeng's direct shareholding ratio has risen to 34%. In this way, Liang Wenfeng's controlling right is more visible — if a financing due - diligence is carried out, DeepSeek's equity structure will appear clearer.

“Not everyone can participate,” investors sighed sincerely. Indeed, after a fierce battle in the Chinese large - model arena, DeepSeek is still very attractive.

Just as the preview version of DeepSeek V4 has finally made its debut, the Pro version and Flash version come standard with a million - token context. The Pro version has up to 1.6 trillion parameters, and the price is quite appealing: For the Pro version, it costs 1 RMB per million token input (cache hit) or 12 RMB (cache miss), and 24 RMB for output. For the Flash version, it is 0.2 RMB, 1 RMB, and 2 RMB respectively.

Meanwhile, the rumor about the adaptation of domestic chips has been confirmed. In the technical report of DeepSeek V4, although it can be seen that the model training part still probably uses NVIDIA chips, Huawei Ascend and NVIDIA are listed side by side on the verification platform. “It is expected that after the mass - market launch and deployment of the Ascend 950 super - node in the second half of the year, the price of the Pro version will also be significantly reduced.”

This move means that DeepSeek has opened a crack in NVIDIA's solid CUDA ecosystem. The imagination behind it is self - evident.

Liang Wenfeng's Determination: The Real Start of the Domestic AI Era

Liang Wenfeng and DeepSeek are on a path against the consensus.

Normally, the clock of a star technology company goes like this: Seize the financing opportunity when it shows its edge, accelerate talent expansion and product iteration, and occupy the market as soon as possible and seek listing. Once this series of actions starts, it is hard to stop.

But every step of DeepSeek is unexpectedly slower.

When DeepSeek R1 was released in early 2025, Liang Wenfeng had almost no competitors. But when he was in the limelight, he rejected all the investors who came knocking. It is precisely today, when the competition is white - hot and there are many competitors, that DeepSeek has released the financing news for the first time. Most of the external discussions attribute it to two points: Funds are needed for R & D. More deeply, DeepSeek needs to give a definite valuation to its internal talents.

Product iteration also comes late. It has been 15 months since the last major update of DeepSeek. After much anticipation, DeepSeek only released the preview version of V4 first, and the long - awaited multi - modality was not updated synchronously. It was not until April 29 that DeepSeek launched the gray - test image - recognition mode, releasing a signal of multi - modality ability.

“Not tempted by praise, not afraid of slander,” this is DeepSeek's attitude. And the feedback from the market seems to verify the idea that things will turn out well if handled slowly.

On the day when DeepSeek V4 was released, domestic AI chips such as Huawei Ascend, Cambricon, Hygon Information, Moore Threads, Muxi Co., Ltd., Kunlunxin, T - Head Zhenwu, and Daysci Technologies completed the adaptation. For a moment, it triggered a rise in the secondary market for domestic chips.

Meanwhile, news shows that the market demand for Huawei's Ascend 950 series of AI chips has soared. The three domestic leading Internet companies, ByteDance, Tencent, and Alibaba, have started negotiations with Huawei on new chip orders.

Thus, an anti - consensus player has triggered industrial resonance with a slow pace — when the underlying chip manufacturers and leading big companies start to evolve around DeepSeek's standards, DeepSeek may have jumped out of the original competitive table.

Just like the saying: Slow is fast. However, not many people really believe it.

This article is from the WeChat official account “Investment World” (ID: pedaily2012), author: Feng Yuchen, published by 36Kr with authorization.