By Feng Yuchen
Chinese AI startup DeepSeek has retained 97% of its core research and engineering staff over the past 15 months — quietly defying a wave of headlines suggesting a talent exodus to bigger tech rivals.
The company, which finally released its long-awaited V4 model in April, is now preparing its first-ever external funding round, according to people familiar with the matter. For founder Liang Wenfeng, who famously turned away investors at the height of DeepSeek’s fame, the move appears less about raising cash than about giving his unusually loyal workforce a clear, high valuation — locking in that loyalty before competitors can prize it away.
A technical report accompanying the release of DeepSeek V4 showed that only 10 out of roughly 270 research and engineering personnel left the company during the model’s development cycle. That puts core R&D attrition at less than 4%, a strikingly low figure for a fiercely competitive sector where poaching is rife.
The data offers a counterpoint to a series of high-profile departures that have dominated Chinese tech headlines. Researchers including Luo Fuli, Wang Bingxuan and Guo Daya have left DeepSeek over the past year for roles at Xiaomi, Tencent and ByteDance, feeding a narrative of instability at one of the country’s most closely watched AI upstarts.
Skills, creativity, passion
But the report suggests those cases are the exception rather than the rule. “If you are chasing short-term goals, hiring experienced people is certainly correct,” Liang said in a rare public comment in 2023. “But from a long-term perspective, basic skills, creativity and passion are more important.”
DeepSeek has historically favoured hiring recent graduates from elite universities such as Tsinghua and Peking University over industry veterans – a strategy that appears to have fostered loyalty even as competitors wave larger pay packets.
The V4 release coincides with DeepSeek’s quiet entry into the world of external fundraising. After months of speculation, the company is now in early-stage talks with potential investors, according to a financial advisory firm familiar with the matter. The round would be DeepSeek’s first since its founding in 2023.
Valuation expectations are already running high. In April, DeepSeek was reported to be seeking a valuation north of $10 billion while local media later reported a pre-investment valuation of 300 billion yuan ($44 billion), although neither figure has been confirmed.
Liang raises his stake
DeepSeek has made no official comment on the fundraising. However, a corporate registration change on April 27 attracted quiet attention in Beijing tech circles: Liang increased his direct stake in the company from 1% to 34% by raising his subscribed capital from 100,000 yuan to 5.1 million yuan. He now controls roughly 84% of DeepSeek directly and indirectly.
“This is not something most people can participate in,” one investor said, adding that financial investors are expected to have only a very small presence in the round. Tencent and Alibaba have held informal discussions with DeepSeek on day-to-day matters, sources say, but substantive funding talks have not yet taken place.
DeepSeek V4 arrived 15 months after its last major model update – an unusually long gap in an industry where speed is often equated with survival. The preview version still lacks native multimodality, a feature widely seen as a missing piece in DeepSeek’s arsenal. A grayscale test for image recognition began only on April 29.
Liang’s unhurried approach has drawn scepticism, but also industry-wide ripple effects. On the same day V4 was released, more than half a dozen domestic AI chipmakers – including Huawei Ascend, Cambricon and Moore Threads – announced they had completed adaptation to the new model.
A contrarian bet
The Huawei Ascend 950 series has since seen a sharp rise in demand, with ByteDance, Tencent and Alibaba reportedly in talks with Huawei for additional chip orders.
By refusing to raise funds at the height of the R1 frenzy and instead delaying both capital and product cycles, Liang has forced the broader ecosystem to align around DeepSeek’s technical standards. “When bottom-layer chipmakers and leading tech giants begin to calibrate and evolve according to DeepSeek’s yardstick,” one industry observer noted, “DeepSeek may have already jumped off the original competitive table.”
Whether that contrarian bet pays off will depend on how DeepSeek navigates its first external funding round – and whether its exceptionally low attrition rate survives the arrival of outside investors.
Source:
PEdaily.cn
