DeepSeek Launches First External Fundraising Round Amid Core Staff Departures

DeepSeek Confirms First External Fundraising

On April 18, multiple venture capital sources confirmed that DeepSeek has begun its first external fundraising round, according to澎湃新闻 (Pail News). The company is targeting a valuation exceeding $10 billion and plans to raise at least $300 million to bolster its capital reserves amid rising costs in the AI competition, according to reports citing informed sources.

Background: Prior Rejection of Commercialization

DeepSeek previously gained industry recognition for rejecting commercialization focus, relying on founder Liang Wenfeng and backing from Phantasm Capital. The company possessed strong technical capabilities in quantitative trading and intelligent finance sectors and was among China’s first large model companies to operate a 10,000-card GPU cluster.

Core Personnel Departures

Despite DeepSeek’s prominence following its viral success during Chinese New Year last year, the company has experienced significant talent losses. According to澎湃新闻 reporting, multiple core researchers have departed since last year, predominantly “post-95s” young scientists:

Multimodal Model Researcher: On April 12, autonomous driving company YuanRong Autonomous Driving publicly confirmed that Ruan Cong, a core contributor to DeepSeek’s multimodal model, has joined as Chief Scientist and will make his first public appearance at the Beijing Auto Show.

First-Generation LLM Author: Wang Bingxuan, core author of DeepSeek’s first large language model, recently announced joining Tencent.

OCR Series Author: Wei Haoran, core author of the DeepSeek-OCR series, departed around Chinese New Year this year but has not publicly disclosed his new employer.

GRPO Algorithm Researcher: On April 16, former DeepSeek core researcher Guo Daya was reported to have joined ByteDance with a reported salary in the hundreds of millions of yuan. According to related disclosures, Guo Daya joined ByteDance’s Seed organization responsible for large model research and development as one of the agent (intelligent agent) direction leads at L8 level. Guo Daya is identified as a major contributor to the GRPO algorithm, which is core to DeepSeek-R1’s reasoning training methodology. On the same day, ByteDance Group Vice President Li Liang responded that the report was inaccurate and that the company has not recently hired employees at near-hundred-million-yuan annual salary levels. However, according to multiple sources confirmed by澎湃新闻, Guo Daya has indeed joined ByteDance.

Deep Learning Researcher: On November 12, former DeepSeek core researcher Luo Fuli publicly announced joining Xiaomi MiMo, stating in a social media post: “Intelligence will eventually transition from language to the physical world. I am at Xiaomi MiMo, working with a group of creative, talented, and genuinely passionate researchers to build this future and pursue the AGI we envision.” According to public information, Luo Fuli graduated from Beijing Normal University’s Computer Science program and completed a master’s degree in computational linguistics at Peking University. After her master’s degree, she joined Alibaba DAMO Academy as a machine intelligence laboratory researcher developing multilingual pre-training model VECO and promoting AliceMind open-source work. In 2022, Luo Fuli joined Phantasm Quantitative (DeepSeek’s parent company) for deep learning work, later serving as a DeepSeek deep learning researcher and participating in the research and development of models including DeepSeek-V2.

Talent Drain Across Multiple Domains

Based on the above information, DeepSeek has experienced core talent losses across multiple domains including foundation large language models (LLM), intelligent agents (Agent), optical character recognition (OCR), and multimodal technologies.

According to industry sources, DeepSeek’s salary and compensation levels are mid-tier in the industry, not the highest. However, headhunters are currently accelerating talent poaching from DeepSeek’s team with 2-3x higher salaries and equity options, accelerating personnel losses.

Platform Updates and V4 Expectations

On April 8, new interface updates were observed on DeepSeek: the input box now displays “Quick Mode” and “Expert Mode” options. According to the webpage display, Quick Mode is suited for daily conversations with immediate responses and supports text recognition from images and files, while Expert Mode excels at complex problems. This marks DeepSeek’s first introduction of layered modes on its official webpage.

These updates have renewed speculation about DeepSeek’s V4 release. Based on external media reports and information from social media and multiple sources, DeepSeek is expected to formally launch V4 in April. According to external expectations, if this V4 release is to replicate last year’s Chinese New Year phenomenon, it will undoubtedly face greater challenges, and personnel losses will inevitably impact the V4 release.