DeepSeek Loses 10 Key Researchers as China's AI Talent War Intensifies

DeepSeek, the Chinese AI lab behind some of the most talked-about open-source models in 2024-2025, has lost at least 10 core researchers and engineers from its ranks — a revelation buried in the technical report for its latest DeepSeek V4 model. The departures underscore a talent war in China's AI sector that is now moving faster than the model releases themselves.

The nearly 300-person author list for DeepSeek V4 quietly flagged 10 names as 'former employees,' including several researchers who were instrumental in building DeepSeek's most important models. Their destinations read like a who's-who of China's AI ecosystem: ByteDance, Tencent, and autonomous driving startup DeepRoute.ai (元戎启行).

Key Takeaways

• 10 core researchers listed in the DeepSeek V4 technical report have already left the company
• Daya Guo, a pivotal DeepSeek researcher, has joined ByteDance's Seed team as an Agent lead at the senior L8 level
• Bingxuan Wang, a core author of DeepSeek's first-generation LLM, has moved to Tencent's Hunyuan team
• Chong Ruan, a multimodal research lead, now serves as Chief Scientist at autonomous driving firm DeepRoute.ai
• ByteDance's Seed team is rapidly becoming China's most aggressive recruiter of top-tier AI talent
• The talent churn suggests China's AI talent pool remains dangerously shallow relative to demand

Who Left DeepSeek — and Where Did They Go?

The departures are not random attrition. These are foundational contributors to DeepSeek's technical breakthroughs.

Daya Guo was one of DeepSeek's most prominent researchers, contributing to core model architecture and training methodology. He has joined ByteDance's Seed team — the company's elite AI research division — where he now leads Agent development efforts. His appointment at the L8 level signals that ByteDance is willing to offer principal-researcher-grade positions and compensation packages to lure top talent.

Bingxuan Wang was a core author on the original DeepSeek large language model, making him one of the architects of the system that put DeepSeek on the map. He has moved to Tencent's Hunyuan team, which is building the tech giant's flagship foundation models.

Chong Ruan, who specialized in multimodal AI at DeepSeek, has taken on the role of Chief Scientist at DeepRoute.ai, an autonomous driving company. This move reflects a broader trend of LLM researchers crossing into robotics and autonomous systems, where multimodal understanding is becoming a critical differentiator.

Haoran Wei, the lead behind DeepSeek's OCR model series, has also departed, though his next destination has not been publicly disclosed. Six other senior contributors — Haowei Zhang, Jun Ran, Junlong Li, Kezhao Huang, Y.Q. Wang, and Zipeng Zhang — are also confirmed to have left.

ByteDance's Seed Team Emerges as China's AI Talent Magnet

ByteDance's Seed team has become what industry observers in China are calling the 'Whampoa Military Academy of AI' — a reference to the famous Chinese military academy that trained a generation of leaders. But in this case, the metaphor cuts both ways: Seed is both producing and aggressively acquiring talent.

The Seed team sits at the heart of ByteDance's AI ambitions. It is responsible for the company's foundation model development, powering products across TikTok, Douyin, and the company's growing suite of enterprise AI tools. The team has been on a hiring spree throughout 2024 and into 2025, pulling researchers from:

• DeepSeek (including Daya Guo and reportedly others)
• Tsinghua University and other elite Chinese research institutions
• Baidu, Alibaba, and other major Chinese tech companies
• Overseas returnees from Google DeepMind, Meta FAIR, and U.S. universities

ByteDance's advantages in this talent war are significant. The company offers competitive compensation that often exceeds what pure-research labs like DeepSeek can provide. It also offers something arguably more attractive to ambitious researchers: scale. With over 1 billion users across its platforms, ByteDance provides an immediate deployment pipeline for AI research that few organizations can match.

Why DeepSeek Is Vulnerable to Poaching

DeepSeek's talent losses highlight a structural vulnerability that many research-focused AI labs face. The company, backed by quantitative trading firm High-Flyer Capital, achieved remarkable results with relatively lean teams and efficient training approaches — its DeepSeek V2 model famously demonstrated that competitive performance could be achieved at a fraction of the cost of Western counterparts.

But that efficiency-first culture creates a paradox. Researchers who prove they can do more with less become extraordinarily valuable acquisition targets for larger, better-funded organizations. The very skills that made DeepSeek's team exceptional — deep expertise in model optimization, training efficiency, and architectural innovation — are precisely what every major AI lab in the world is desperate to hire.

DeepSeek also faces a compensation gap. While the lab reportedly pays well by Chinese research standards, it cannot easily compete with the total compensation packages offered by ByteDance, Tencent, or Alibaba, which combine base salary, stock options, and bonuses that can reach into the millions of dollars for senior researchers.

Furthermore, the departure of Luo Fuli, a core developer on DeepSeek V2, adds to the pattern. When key researchers leave in clusters, it can trigger a cascade effect — remaining team members may feel uncertain about the lab's stability, accelerating further departures.

China's AI Talent Pool Remains Shallow

The aggressive poaching cycle reflects a fundamental supply-demand imbalance. China has a growing number of AI companies and well-funded AI divisions, but the pool of researchers with hands-on experience building frontier-class large language models remains remarkably small.

Compared to the United States, where talent circulates among dozens of well-established AI labs — OpenAI, Anthropic, Google DeepMind, Meta FAIR, xAI, Cohere, and numerous startups — China's ecosystem is more concentrated. A relatively small number of individuals have been directly involved in training models at the frontier, and they are being fought over with increasing intensity.

This dynamic mirrors what happened in the U.S. market between 2022 and 2024, when researchers cycled rapidly between OpenAI, Google, Anthropic, and various startups. The difference is that China's talent pool is even more constrained, and the stakes for individual companies are correspondingly higher.

Key factors driving the talent war include:

• Exploding demand: Every major Chinese tech company now treats foundation model development as a strategic priority
• Limited supply: Fewer than a few thousand researchers in China have direct experience with frontier model training
• Geopolitical pressure: U.S. chip export restrictions have made software and algorithmic innovation even more critical, increasing the value of top researchers
• Startup proliferation: New AI startups like Moonshot AI, MiniMax, Zhipu AI, and 01.AI are all competing for the same talent pool
• Cross-sector demand: Autonomous driving, robotics, and healthcare AI companies are now recruiting LLM researchers

What This Means for the Global AI Landscape

For Western observers, China's AI talent war carries several important implications.

First, it suggests that DeepSeek's competitive advantage may be more fragile than its impressive technical reports suggest. If the lab continues to lose core contributors at this rate, its ability to maintain its rapid model release cadence could be compromised — though the company has historically shown resilience and an ability to develop new talent internally.

Second, ByteDance's aggressive talent acquisition signals that the company is preparing for a major push in AI product development. The Seed team's growing roster of elite researchers positions ByteDance as potentially the strongest AI player in China, with implications for global competition in AI-powered content, search, and enterprise services.

Third, the talent churn creates knowledge diffusion across the Chinese AI ecosystem. When researchers move between organizations, they carry architectural insights, training techniques, and institutional knowledge. This accelerates the overall pace of AI development in China, even as it may weaken individual organizations.

Looking Ahead: Will the Churn Slow Down?

The current pace of AI talent movement in China appears unsustainable. Companies are paying ever-higher premiums for experienced researchers, and the frequent job-hopping can disrupt long-term research programs.

Several potential stabilizing forces could emerge. Companies may increasingly turn to non-compete agreements and retention bonuses to lock in key talent. Research labs like DeepSeek may offer equity stakes or profit-sharing arrangements that make departure more costly. And the growing output of Chinese university AI programs will gradually expand the talent pool, though it will take years for new graduates to reach the expertise level of the researchers currently being poached.

For now, however, the talent war shows no signs of cooling. As one Chinese AI industry commentator noted, 'The speed of talent movement between companies is outpacing the speed of model iteration — and that is saying something in an industry that releases new models every few months.'

The DeepSeek V4 author list, with its 10 annotated departures, may become a defining artifact of this era — a technical document that inadvertently tells the story of an industry where human capital, not just compute, remains the scarcest and most contested resource.