THE HOOVER INSTITUTION TPA WHITE PAPER SERIES

A Deep Peek into DeepSeek AI’s Talent

and Implications for US Innovation

By Amy Zegart and Emerson Johnston

APRIL 21, 2025

A collaboration with
The Stanford Institute for Human-Centered AI
 A Deep Peek into DeepSeek AI’s Talent and
Implications for US Innovation

By Amy Zegart and Emerson Johnston

Chinese startup DeepSeek AI has upended conventional wisdom about artificial

intelligence (AI) innovation. Released in January 2025, the company’s R1 language
model and V3 general-purpose large language model (LLM) sent tremors through
markets and challenged assumptions about American technological superiority in
frontier AI development.1 Although DeepSeek AI’s claims that its V3 model was trained
for just $6 million have been widely disputed (experts estimate the true compute costs
are closer to half a billion dollars, and DeepSeek AI itself says the cost was just for the
final training run), the R1 model built on top of V3 demonstrated unprecedented
reasoning capabilities and technical achievements that surpassed previous benchmarks
set by US companies.

Beneath DeepSeek’s technical achievements lies a more consequential story: the

shifting patterns of global AI talent that made the company’s breakthroughs possible.
This paper examines the educational backgrounds, career paths, and international
mobility of more than 200 researchers who authored DeepSeek’s five foundational
papers from January 2024 to February 2025. These five papers constitute the corpus of
the company’s openly available research papers since its founding in 2023.

We find striking evidence that China has developed a robust pipeline of homegrown
talent. Nearly all of the researchers behind DeepSeek’s five papers were educated or
trained in China. More than half of them never left China for schooling or work,
demonstrating the country’s growing capacity to develop world-class AI talent through
an entirely domestic pipeline. And while nearly a quarter of DeepSeek researchers
gained some experience at US institutions during their careers, most returned to China,
creating a one-way knowledge transfer that benefits China’s AI ecosystem.

These talent patterns represent a fundamental challenge to US technological

leadership that export controls and computing investments alone cannot address.
DeepSeek is an early-warning indicator about the essential role that human capital—
not just hardware or algorithms—plays in geopolitics, and how America’s talent
advantage is eroding.

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 2

 THE HOOVER INSTITUTION TPA WHITE PAPER SERIES

Methodology

DeepSeek AI, a Chinese AI research company focused on “cost-efficient, high-

performance language models,” released five papers on Cornell University’s arXiv.org
manuscript archive between 2024 and 2025.2 A total of 223 authors were credited
across the five papers. We were able to conduct a comprehensive review of 211 of
them.3 Using data from the OpenAlex research catalog collected in February 2025, we
collected detailed author profiles (publication records, citation metrics, institutional
affiliations dating back to 1989) and comprehensive institutional data (geographical
location, organization type, research output metrics), paying special attention to
tracking changes over time.4 Then, through custom Python scripts for data collection
and analysis, we mapped each researcher’s complete institutional history, revealing
previously undetected patterns of cross-border movement. While traditional analyses
often rely on static snapshots of talent at a particular point in time, our approach
allowed us to quantify not just where talent is today, but how it has flowed between
countries over time—particularly between China and the United States—capturing the
“reverse brain drain” cases that represent strategic knowledge-transfer mechanisms.

DeepSeek’s Talent Infrastructure Across Five Papers

A total of 223 people were listed as contributors to any of DeepSeek’s five papers (see
fig. 1). Our analysis finds that 31 researchers (or just under 14 percent of the total
author pool) contributed to all five papers—what we refer to as the “Key Team.”5
Another 50 authors worked on four papers, 64 contributed to three papers, 55 were
listed on two papers, and 22 researchers contributed to just one paper.

Figure 1 – Authorship Distribution Across DeepSeek Papers

Distribution of 222 authors across the five DeepSeek AI papers, with 31 authors (14%) contributing to all five papers

Source: All data from OpenAlex.

3
 As table 1 illustrates (see appendix B), it appears that DeepSeek used a shifting
categorization of talent across the five papers. In Paper 1 (DeepSeek LLM), the
reported contributor labels were organizational rather than role-based, with 53
individuals categorized into Business Team (8), Compliance Team (7), Data Annotation
Team (36), and Design Team (2). Notably, none of these labeled contributors were
credited as authors on the paper itself, which officially listed 86 authors. However, 40 of
those contributors were later credited as authors in at least one subsequent DeepSeek
paper—which is why they are captured here. This discrepancy may suggest that Paper
1’s contributor list reflected a broader pool of internal collaborators—many of whom
were not formally recognized at the time but went on to receive authorship credit as
the project evolved.

Papers 2 and 4 appear to have transitioned to more functionally descriptive categories

that closely resembled internal team structures. Paper 2 introduced hybrid contributor
tags such as “Business & Compliance,” “Data Annotation,” and “Research &
Engineering.” Among the 156 total contributors, the vast majority (105) were classified
under Research & Engineering, followed by 31 in Data Annotation and 18 in Business &
Compliance. Notably, 2 contributors—Shengfeng Ye and Yanhong Xu—were listed in
more than one category: Ye appeared in both Research & Engineering and Business &
Compliance, while Xu was credited under Research & Engineering and Data
Annotation, likely reflecting overlapping responsibilities within the organization. Paper
4 (DeepSeek V3), which had 197 listed authors, followed the same categorization
structure: Research & Engineering (148), Data Annotation (30), and Business &
Compliance (17). Again, Ye and Xu were the only contributors assigned to two
categories—Ye in Research & Engineering and Business & Compliance; Xu in Research
& Engineering and Data Annotation. This schema appears to capture the backbone of
DeepSeek’s technical efforts. Within this structure, nearly all members of the 31-person
Key Team were designated as Research & Engineering contributors, with one notable
exception—Yanhong Xu, who, as noted, also held a Data Annotation role in Papers 2
and 4.

Papers 3 and 5 introduced a different delineation between levels of contribution

through a binary categorization: Contributor and Core Contributor. This shift may
indicate a formal recognition of hierarchical status within the research group. In Paper
3, just 4 of the 39 contributors were labeled as Core Contributors. Similarly, Paper 5—
the company’s internationally watched R1 reasoning model—designated 18 Core
Contributors out of 194 total. In both cases, Core Contributors made up roughly 10
percent of the total contributor base, suggesting a carefully curated leadership tier.
Notably, all four Core Contributors from Paper 3—Daya Guo, Dejian Yang, Qihao Zhu,

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 4

 THE HOOVER INSTITUTION TPA WHITE PAPER SERIES

and Zhihong Shao—were also credited as Core Contributors in Paper 5, and all four
contributed to every one of the five DeepSeek papers, likely signaling their central,
long-term influence on the DeepSeek project.

DeepSeek Researcher Citation Metrics: Not So Green After All

The prevailing narrative has been that DeepSeek succeeded with younger, less
experienced researchers. Citation metrics, however, suggest that DeepSeek’s talent
was not so green after all.

While the structure of DeepSeek’s collaboration shows clear differentiation in

participation levels, there is also meaningful variation in scholarly experience across
those tiers. Among the set of 211 contributors for whom we were able to pull data, the
average researcher has published sixty-one works and received just over one thousand
citations, with an h-index of 10.8 and i10-index of just over 19.6 It is worth noting that
these averages mask a bimodal distribution: Many researchers have modest academic
footprints, but a concentrated group ranks far higher in output and impact. The median
citation count (249), h-index (7), and i10-index (5) for this group underscore this internal
variation.

Notably, the 31 Key Team researchers who contributed to all five papers stand out
sharply. The group averages 1,554 citations per author, with a median of 501, and a
mean h-index of 13.5 and i10-index of 25.5. Median values—an h-index of 10 and an
i10-index of 11—further indicate consistent impact across the Key Team, not just a few
outliers. These metrics provide additional evidence that the DeepSeek Key Team
consists of researchers with already credible academic track records.

This academic strength becomes even more apparent when compared to a peer group
from one of the world’s leading AI labs. According to data from the OpenAI o1 system
card (arXiv:2412.16720), the team of 265 authors listed on that release had an average
citation count of 4,403 but a median of just 338, indicating a steep drop-off beyond a
few highly cited individuals. Further, the group’s median h-index was only 6 and the
i10-index 4, reflecting more limited consistency of impact across the full group.

In contrast, both DeepSeek’s full author pool and its Key Team exhibit greater balance
between average and median performance—suggesting not only strength at the top,
but also less variation across contributors compared to the OpenAI team. These
patterns may indicate a more evenly distributed base of academic experience, rather
than one overly reliant on a handful of standout figures. DeepSeek’s research engine

5
 appears not only deep but wide—an organizational trait that may prove especially
important as competition in foundation model development intensifies.

Taken together, these comparisons challenge the media narrative that DeepSeek’s
rapid ascent was driven by “untested” or inexperienced researchers. While OpenAI
continues to receive global recognition, many of DeepSeek’s central contributors—at
least by traditional bibliometric standards—were better published, more consistently
cited, and arguably more academically established at the time of their breakthrough.7

A Longitudinal View of Institutional Affiliations: China’s Dominant Position

Looking longitudinally at the 201 DeepSeek authors with known affiliation data, we find
that more than half (n=111) have been trained and affiliated exclusively at Chinese
institutions throughout their careers—evidence of China’s growing capacity to develop
world-class AI talent domestically without relying on Western expertise. And the vast
majority of DeepSeek authors—98 percent (n=197)—have held at least one past or
current affiliation with a Chinese institution.

Four authors appear to have not studied, trained, or worked in China at all. Their
academic and professional roots spanned a range of global institutions: Erhang Li was
trained in the United Kingdom and the United States and is affiliated with Intel UK; Y.
Q. Wang studied in Germany at Johannes Gutenberg University Mainz; Yuduan Wang
received education in Singapore at the National University of Singapore; and Panpan
Huang studied in the United States at Purdue University. While these individuals
represent exceptions within the broader DeepSeek ecosystem, they highlight the
international reach of the global AI research community. Still, their small number
underscores how uncommon this path is among DeepSeek contributors—further
reinforcing the observation that China’s domestic pipeline is now capable of producing
world-class AI researchers largely on its own.

We found that only a quarter of DeepSeek researchers (24.3 percent, n=49) have ever
held an academic or professional affiliation with a US institution—further illustrating the
limited role American institutions have played in shaping this cohort.

China Dominates the 2025 Snapshot of Institutional Affiliations, Too

As figure 2 shows, 171 of the 201 DeepSeek authors with known affiliation data were
affiliated with Chinese institutions in 2025 (the most current year available).8

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 6

 THE HOOVER INSTITUTION TPA WHITE PAPER SERIES

Figure 2 – Geographic Distribution of Current Institutional Affiliations

Current geographic distribution of 201 DeekSeek AI researchers with known affiliations

Source: All data from OpenAlex.

Just 7 percent (n=15) of researchers currently hold US-based affiliations. These include
positions at prominent research universities (such as Stony Brook University, University
of North Texas, and the University of California, San Francisco), medical institutions
(such as Boston Children’s Hospital), and tech or biotech companies including Google,
Otsuka, and Health First. The remaining researchers are spread across a small set of
other countries, including Australia, Canada, the United Kingdom, and Singapore, with
single cases in Germany, Ireland, Panama, Poland, and Taiwan. This geographic
consolidation around China further reinforces the central role of its domestic
institutions—not just as training grounds, but as long-term professional destinations for
AI talent.

The Central Role of the Chinese Academy of Sciences

The broader institutional landscape supporting DeepSeek’s development reflects the

full career trajectories of its researchers, encompassing all known affiliations across
time. In total, the 211 analyzed authors were connected to 499 unique institutions
globally, with Chinese institutions accounting for 368 (74 percent) of them (see fig. 3).
The network is predominantly anchored in academia, with universities and research
institutions forming the backbone, but it also features some training from private
companies (n=17), government institutions (n=12), and nonprofit organizations (n=9).

7
 Figure 3 – Top 10 Institutions by Researcher Count

Source: All data from OpenAlex.

Within this institutional landscape, the Chinese Academy of Sciences (CAS) emerges as
the strategic center of gravity. While directly hosting only 18 authors, CAS
encompasses a total of 53 researchers when accounting for its network of 153 affiliated
institutions. This extensive institutional reach—where a “child institution” refers to an
organization with a subsidiary relationship (as defined by OpenAlex) to CAS as its
parent organization, including research institutes, laboratories, and specialized
centers—combined with remarkable research impact metrics (over 840,000 works and
23.7 million citations), positions CAS as the dominant player in this ecosystem.9 Peking
University comes in second with 21 total affiliations, but it leads in direct affiliations
with 20 researchers. Tsinghua University follows with 16 authors, then Sun Yat-sen
University and Nanjing University with 10 authors each. This distribution reveals how
China has leveraged its institutional infrastructure to support AI development, with a
network centered around CAS but distributed across multiple prestigious universities.
The concentration of talent within this network of Chinese institutions has created a
fertile environment for AI innovation that challenges the US advantage in institutional
resources.

The US-China Talent Pipeline: Challenging American Assumptions

Of the 49 DeepSeek researchers who had US affiliations at some point during their
careers, 63.3 percent (n=31) spent just one year in the United States—long enough to
gain exposure to top-tier research environments, but not long enough to establish
enduring ties. Another 18.4 percent (n=9) remained for two to four years, and 18.4
percent (n=9) stayed five years or longer, often across multiple institutions. This latter
group includes some of the most influential researchers in the cohort, such as Minghua
Zhang, who accumulated affiliations at State University of New York and Stony Brook

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 8

 THE HOOVER INSTITUTION TPA WHITE PAPER SERIES

University spanning over a decade; Zhenda Xie, who spent eight years across UCLA
and Optica; and B. Zhang, whose recurring ties with the University of Southern
California from 2007 to 2022 preceded his return to Peking University. These 9 long-
stay researchers are not statistical outliers—they averaged 4,541 citations, held a
median h-index of 25, and had a median i10-index of 40. Despite this deep academic
integration, only 3 of the 9 currently remain affiliated with US institutions, further
underscoring how the US research ecosystem served as a powerful incubator of talent
that ultimately advanced China’s AI leadership (see fig. 4).

Figure 4 – US Experience Duration

US Experience Duration for the 49 DeepSeek Researchers with US Affiliations

Source: All data from OpenAlex.

Notably, the institutional diversity of US experience among DeepSeek researchers is

significant. The 49 individuals with US affiliations were connected to 65 different
institutions across 26 states, including public universities, private colleges, medical
centers, nonprofit organizations, and technology companies. While no single institution
accounted for more than three researchers, several—including the University of
Southern California, Stanford University, and New York University—had multiple
affiliations. This distribution spans the full geographic breadth of the United States,
with clusters visible in in key innovation hubs: the Bay Area and Southern California, the
Boston-to-DC corridor, and research-heavy regions of Texas and the Midwest (see fig.
5). Importantly, rather than concentrating within a small number of elite campuses,
these researchers engaged with a wide cross-section of the American research
ecosystem. This breadth may have facilitated broader exposure to US scientific and
technological practices. It also meant that no single institution had good visibility into
the scale of the international AI knowledge exchange taking place.

9
 Figure 5 – Geographic Distribution of US Institutions Affiliated with DeepSeek Researchers

Source: All data from OpenAlex.

More telling than location or duration is direction. Among the 49 DeepSeek

researchers with US affiliations, only a small share followed the linear trajectory of
moving from China to the United States and remaining in the US (see figs. 6a and 6b).
Instead, our data shows that the dominant mobility pattern is cyclical, multinational,
and strategically adaptive. As shown in figure 6b, almost 40 percent (n=19) of these
researchers began their careers in China, traveled abroad—including to the United
States—and ultimately returned to China. Researchers such as Xuan Lu, Xiaodong Liu,
and Shiyu Wang exemplify this classic “study-abroad-and-return” model: China → USA
→ China. Their careers reflect a traditional, state-aligned mobility model where US
training is used to strengthen domestic capabilities.

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 10

 THE HOOVER INSTITUTION TPA WHITE PAPER SERIES

Figure 6a – US Retention Rate

US Retention Data for the 49 DeepSeek Researchers with US Affiliations

Figure 6b – US Retention Rate (simplified)

US Retention Rate for the 49 DeepSeek Researchers with US Affiliations

Source: All data from OpenAlex.

A second, more complex group includes researchers such as Wenfeng Liang, Minghua
Zhang, and Zhiyu Wu, whose careers span multiple transits between China and the
United States (e.g., China → USA → China → USA → China). These researchers don’t
simply return—they circulate, developing global networks and embedding themselves
in both ecosystems. This pattern of bidirectional exchange accounts for 12.2 percent
(n=6). Of these, 4 currently list US institutions as their most recent affiliation, while 2 are
affiliated with Chinese institutions. While it is difficult to determine intent or long-term
plans from affiliation data alone, these cases illustrate how cross-border mobility can
strengthen China’s AI ecosystem without necessarily requiring permanent US retention.

11
 Other researchers such as Daya Guo, Guanting Chen, and Yicheng Wu take even more
global paths—passing through institutions in the United Kingdom, Singapore, Saudi
Arabia, Taiwan, or Australia. These trajectories (e.g., Taiwan → China → Australia →
USA) illustrate the rising influence of multinational knowledge acquisition, with the
United States serving as just one of many strategic destinations in a broader global
loop.

Notably, only 14.2 percent (n=7) of the DeepSeek cohort (e.g., Ruiqi Ge, Peiyi Wang,
Bingxuan Wang) followed a China → USA → Stayed path—remaining in the United
States after initial training (see fig. 6a). While still significant, this group is no longer the
default or dominant outcome. Even among researchers who began in the United States
(e.g., B. Zhang, Ruoyu Zhang), many ultimately relocated to China, with 22.4 percent
(n=11) falling into the “Started in the USA, Ended Up in China” or “Started in the USA,
Traveled, Ended Up in China” categories.

Finally, a small but illustrative set of researchers defies simple classification. Figures
such as Kuai Yu (USA → Netherlands → Singapore → China) or Zhen Zhang (USA →
China → Hong Kong → USA) reflect the complexity of today’s scientific mobility. These
researchers—counted within “Started in USA, Traveled, Ended Up in China” (6.1
percent, n=3) or “Started in USA, Traveled, Ended Up in USA” (4.1 percent, n=2)—
reveal how transnational scientific careers are increasingly nonlinear and dynamic.

Taken together, these patterns reveal important features of global AI talent flows. The
United States remains a vital node in international research training—but it is not the
fulcrum or the end point. Most of DeepSeek’s researchers are not being trained in the
United States, and those who are trained here are not retained. Instead, they are
passing through. These findings suggest that American institutions are serving as
steppingstones, equipping elite researchers with high-impact skills, connections, and
credentials that are ultimately reinvested into China’s AI ecosystem. Importantly, the 49
DeepSeek researchers with US affiliations at some point in their careers were among
the most academically accomplished in the entire research cohort, averaging 2,168
citations (median 565), with a mean h-index of 17 and i10-index of 34—figures
significantly higher than those for the broader DeepSeek author pool. These are not
peripheral actors, but central contributors to one of China’s most advanced AI efforts.

For US policymakers, our DeepSeek talent analysis suggests it is high time to reassess
long-standing assumptions that the world’s best and brightest naturally want to study
and stay in the United States. Attracting and permanently retaining the world’s best
minds—once a cornerstone of American technological dominance—appears
increasingly misaligned with twenty-first-century educational realities. DeepSeek is, at

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 12

 THE HOOVER INSTITUTION TPA WHITE PAPER SERIES

its core, a story of homegrown capacity: Half of its researchers have never left China,
the overwhelming majority have deep institutional ties to China, and even many who
trained in the United States ultimately returned to China—potentially advancing
China’s position in the global AI race.

A Closer Look at the Key Team and Where They Trained

Of 31 Key Team authors, 28 had available institutional affiliation data on OpenAlex.

Half of them (n=14) have spent part of their careers at institutions outside of China.
These globally mobile researchers often followed targeted international pathways:
initial training at elite Chinese universities followed by graduate study, postdoctoral
work, or research appointments abroad—typically in the United States, the United
Kingdom, Australia, or other key AI hubs—before returning to China.

Notable examples include Daya Guo (China → UK → China → USA → UK → China),

who spent time at both Rensselaer Polytechnic Institute and Microsoft Research (United
Kingdom); Jiashi Li (China → Japan → China → USA → China), affiliated with the Hoshi
University in Japan and later the University of California, Santa Barbara; and Dejian
Yang (China → UK → Australia), who held positions at the Pharmaron (United Kingdom)
and the University of Technology Sydney. Others, such as Zhenda Xie and Wenfeng
Liang, show repeated, multidirectional mobility between the United States and China,
suggesting enduring cross-border collaboration.

Of the internationally experienced Key Team members, 8 had US affiliations, including

Peiyi Wang (Boston College), Qihao Zhu (Carnegie Mellon University), and Zhihong
Shao (University of Michigan–Ann Arbor). Others had connections to institutions in
Canada (Liyue Zhang), Singapore (Qihao Zhu), Bangladesh (Kai Dong), and South
Korea (Junxiao Song). This distribution reflects a deliberate emphasis on experience in
countries that are global leaders in AI research and higher education.

These patterns suggest a sophisticated approach to human capital development that

treats international experience not as “brain drain” but as strategic national
investment—sending promising researchers abroad to acquire cutting-edge
knowledge and methodologies before returning to apply these assets to China’s
technological advancement.

13
 Geopolitical Implications

The talent patterns revealed in our analysis have significant geopolitical implications.
For centuries, the sources of national power have stemmed from tangible assets—such
as territory that could be conquered, populations that could be taxed or conscripted,
goods that could be embargoed, militaries that could be deployed. Those tangible
sources of national power still matter, but in the technology age, power also derives
from intangible assets such as data, technology, and knowledge inside people’s heads.
Knowledge power has never been more important for economic and geopolitical
competition; it is the ultimate portable weapon.

These findings challenge a long-held belief that the United States will always attract the
world’s best talent. In reality, however, top global talent has options. DeepSeek’s talent
story suggests that the United States cannot assume a permanent talent lead. Instead,
the nation needs to compete much more aggressively to attract, welcome, and retain
the world’s best and brightest while urgently growing domestic capabilities by
improving K‒12 STEM (science, technology, engineering, mathematics) education at
home.

Ultimately, DeepSeek AI represents more than just another advance in language model
technology. It reveals talent patterns that challenge long-held US assumptions about
innovation advantage. Our analysis of DeepSeek’s research network suggests that
conventional wisdom about US dominance in talent development and retention may
no longer hold true, with significant implications for future technological competition.

ABOUT THE AUTHORS

Amy Zegart is the Morris Arnold and Nona Jean Cox Senior Fellow at the Hoover
Institution, director of Hoover’s Technology Policy Accelerator, and a senior fellow and
associate director at Stanford University’s Institute for Human-Centered AI.

Emerson Johnston is a second-year master’s student in international policy and a

Knight-Hennessy Scholar at Stanford University.

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 14

 THE HOOVER INSTITUTION TPA WHITE PAPER SERIES

Appendix A: DeepSeek Research Papers (2024–2025)

The following five papers released by DeepSeek AI between January 2024 and
February 2025 formed the basis for our institutional and author-trajectory analysis:

1. DeepSeek LLM: Scaling Open-Source Language Models with Longtermism

arXiv:2401.02954 ‒— January 2024
Focuses on scaling laws for open-source LLMs in 7B and 67B configurations,
contributing insights into training efficiency.

Full Abstract: The rapid development of open-source large language models

(LLMs) has been truly remarkable. However, the scaling law described in
previous literature presents varying conclusions, which casts a dark cloud over
scaling of LLMs. We delve into the study of scaling laws and present our
distinctive findings that facilitate scaling of large-scale models in two commonly
used open-source configurations, 7B and 67B. Guided by the scaling laws, we
introduce DeepSeek LLM, a project dedicated to advancing open-source
language models with a long-term perspective. To support the pretraining
phase, we have developed a dataset that currently consists of two trillion tokens
and is continuously expanding. We further conduct supervised fine-tuning (SFT)
and direct preference optimization (DPO) on DeepSeek LLM Base models,
resulting in the creation of DeepSeek Chat models. Our evaluation results
demonstrate that DeepSeek LLM 67B surpasses Llama-2 70B on various
benchmarks, particularly in the domains of code, mathematics, and reasoning.
Furthermore, open-ended evaluations reveal that DeepSeek LLM 67B Chat
exhibits superior performance compared to GPT-3.5.

2. DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts

Language Model
arXiv:2405.04434 ‒ May 2024
Introduces a 236B parameter Mixture-of-Experts (MoE) model with a focus on
cost-effective training and inference using novel architectural choices such as
Multi-head Latent Attention (MLA).

Full Abstract: We present DeepSeek-V2, a strong Mixture-of-Experts language

model characterized by economical training and efficient inference. It comprises
236B total parameters, of which 21B are activated for each token, and supports
a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures
including Multi-head Latent Attention and DeepSeekMoE. MLA guarantees
efficient inference through significantly compressing the Key-Value (KV) cache

15
 into a latent vector, while DeepSeekMoE enables training strong models at an
economical cost through sparse computation. Compared with DeepSeek 67B,
DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves
42.5 percent of training costs, reduces the KV cache by 93.3 percent, and
boosts the maximum generation throughput to 5.76 times. We pretrain
DeepSeek-V2 on a high-quality and multisource corpus consisting of 8.1T
tokens, and further perform supervised fine-tuning and reinforcement learning
(RL) to fully unlock its potential. Evaluation results show that, even with only 21B
activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier
performance among open-source models.

3. DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code

Intelligence
arXiv:2406.11931 ‒ June 2024
A code-specialized MoE model achieving performance comparable to GPT-4
Turbo, emphasizing large-scale continued pretraining.

Full Abstract: We present DeepSeek-Coder-V2, an open-source Mixture-of-

Experts code language model that achieves performance comparable to GPT4-
Turbo in code-specific tasks. Specifically, DeepSeek-Coder-V2 is further
pretrained from an intermediate checkpoint of DeepSeek-V2 with an additional
six trillion tokens. Through this continued pretraining, DeepSeek-Coder-V2
substantially enhances the coding and mathematical reasoning capabilities of
DeepSeek-V2, while maintaining comparable performance in general language
tasks. Compared to DeepSeek-Coder-33B, DeepSeek-Coder-V2 demonstrates
significant advancements in various aspects of code-related tasks as well as
reasoning and general capabilities. Additionally, DeepSeek-Coder-V2 expands
its support for programming languages from 86 to 338, while extending the
context length from 16K to 128K. In standard benchmark evaluations,
DeepSeek-Coder-V2 achieves superior performance compared to closed-source
models such as GPT4-Turbo, Claude 3 Opus, and Gemini 1.5 Pro in coding and
math benchmarks.

4. DeepSeek-V3 Technical Report

arXiv:2412.19437 ‒ December 2024
Advances the DeepSeek MoE line with 671B total parameters and pioneering
loss-free loading techniques to enhance inference efficiency.

Full Abstract: We present DeepSeek-V3, a strong Mixture-of-Experts language

model with 671B total parameters with 37B activated for each token. To achieve

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 16

 THE HOOVER INSTITUTION TPA WHITE PAPER SERIES

efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head

Latent Attention and DeepSeekMoE architectures, which were thoroughly
validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-
loss-free strategy for load balancing and sets a multitoken prediction training
objective for stronger performance. We pretrain DeepSeek-V3 on 14.8 trillion
diverse and high-quality tokens, followed by supervised fine-tuning and
reinforcement learning stages to fully harness its capabilities. Comprehensive
evaluations reveal that DeepSeek-V3 outperforms other open-source models
and achieves performance comparable to leading closed-source models.
Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800
GPU hours for its full training. In addition, its training process is remarkably
stable. Throughout the entire training process, we did not experience any
irrecoverable loss spikes or perform any rollbacks. The model checkpoints are
available at https://github.com/deepseek-ai/DeepSeek-V3.

5. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement

Learning
arXiv:2501.12948 ‒ January 2025
The flagship reasoning-focused model trained via large-scale RL without
supervised fine-tuning. Widely seen as a breakthrough in emergent reasoning
behaviors and the focal point of DeepSeek’s impact.

Full Abstract: We introduce our first-generation reasoning models, DeepSeek-

R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale
reinforcement learning without supervised fine-tuning as a preliminary step,
demonstrates remarkable reasoning capabilities. Through RL, DeepSeek-R1-
Zero naturally emerges with numerous powerful and intriguing reasoning
behaviors. However, it encounters challenges such as poor readability and
language mixing. To address these issues and further enhance reasoning
performance, we introduce DeepSeek-R1, which incorporates multistage training
and cold-start data before RL. DeepSeek-R1 achieves performance comparable
to OpenAI o1-1217 on reasoning tasks. To support the research community, we
open-source DeepSeek-R1-Zero, DeepSeek-R1, and six dense models (1.5B, 7B,
8B, 14B, 32B, 70B) distilled from DeepSeek-R1 based on Qwen and Llama.

17
 Appendix B: Tables

Table 1: Author Count and Contributor Roles for DeepSeek Publications (2024–25)

Publication Date No. of Contributor Categories

Authors* (no. of contributors)

Paper 1: DeepSeek LLM January 2024 86 Business Team (8), Compliance Team (7),
Data Annotation Team (36), Design Team
(2)**

Paper 2: DeepSeek V2 May 2024 156 Research & Engineering (105), Data
Annotation (31), Business & Compliance
(18), Mixed Roles (2: Data Annotation +
R&E, Business & Compliance + R&E)

Paper 3: DeepSeek VCoder 2 June 2024 39 Core Contributor (4), Contributor (35)

Paper 4: DeepSeek V3 December 2024 197 Research & Engineering (148), Data
Annotation (30), Business & Compliance
(17), Mixed Roles (2: Data Annotation +
R&E, Business & Compliance + R&E)

Paper 5: DeepSeek R1 January 2025 200† Core Contributor (18), Contributor (176) –
This total number (194) reflects the no. of
authors from the PDF.

* This number reflects unique authors listed for each paper, consolidating names across both the ArXiv
and PDF versions where applicable. Discrepancies between sources are noted:
● Paper 2 and Paper 4 PDFs each contain two duplicate names (Shengfeng Ye and Yanhong Xu),
due to those individuals being listed in multiple contributor categories.
● Paper 5:
o The PDF includes one duplicate name (Shengfeng Ye).
o The ArXiv includes two duplicate names (Shengfeng Ye and Yanhong Xu).
o Nine authors appear on only one version (either ArXiv or PDF), including:
▪ On ArXiv but not on PDF: Chenyu Zhang, Han Bao, Haocheng Wang, Huajian Xin, Jiawei
Wang
▪ On PDF but not on ArXiv: Jinhao Tu, Kaichao You, Mingxu Zhou, Wanjia Zhao
** The contributor categories listed for Paper 1 reflect a separate contributor pool that was not credited
as authors on that paper. The numbers shown in the chart represent the total number of individuals in
each category at that time.
† Paper 5 showed discrepancies in authorship counts: The PDF version originally listed 195 authors, but
one author (Shengfeng Ye) was listed twice, resulting in 194 unique names. The ArXiv entry listed 197
authors. When combining both lists and removing duplicates, the total came to 201 unique authors.

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 18

 THE HOOVER INSTITUTION TPA WHITE PAPER SERIES

Table 2: List of Key Team Researchers

The list below includes the 31 individuals who are credited as authors on all five DeepSeek AI
papers. An asterisk (*) indicates those identified as core contributors in the fifth paper.

1. Bingxuan Wang 12. Jiashi Li 23. Xiao Bi*

2. Chenggang Zhao 13. Junxiao Song 24. Xin Xie

3. Chengqi Deng 14. Kai Dong 25. Yanhong Xu

4. Chong Ruan 15. Kang Guan 26. Yaohui Wang

5. Damai Dai* 16. Liyue Zhang 27. Yishi Piao

6. Daya Guo* 17. Peiyi Wang 28. Yuxiang You

7. Dejian Yang 18. Qihao Zhu 29. Zhenda Xie

8. Deli Chen 19. Qiushi Du 30. Zhewen Hao

9. Fuli Luo 20. Shirong Ma 31. Zhihong Shao

10. Hanwei Xu 21. Wenfeng Liang

11. Huazuo Gao 22. Wenjun Gao

19
 Table 3: Scholarly Output and Citation Metrics of DeepSeek and OpenAI Research
Teams

Works Count Cited by Count h-Index i10-Index

All DeepSeek Average 61.057 1,059.218 10.791 19.166

Authors
Median 24.000 249.000 7.000 5.000

DeepSeek Core Average 70.806 1,554.258 13.548 25.548

Group
Median 51.000 501.000 10.000 11.000

DeepSeek US- Average 101.286 2,200.286 17.122 34.265

Affiliated
Authors Median 57.000 565.000 12.000 14.000

OpenAI o1 Average 58.951 4,402.917 12.109 24.955

Authors
Median 16.000 338.000 6.000 4.000

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 20

Table 4: US Institutions Affiliated with DeepSeek Researchers

The following list includes US-based academic, research, medical, and industry
institutions where DeepSeek authors have held prior or current affiliations. This
includes both educational and professional roles. Asterisk (*) indicates a current
affiliation based on the most recent OpenAlex data.

No. of Organization City Country

Authors

3 University of Southern California* Los Angeles United States

2 Auburn University Auburn United States

2 New York University New York United States

2 Stanford University Stanford United States

2 University of California, Santa Barbara Santa Barbara United States

2 University of North Texas* Denton United States

1 National Clinical Research Richmond United States

1 University Research Co. (United States) Bethesda United States

1 Electric Power Research Institute Palo Alto United States

1 Johns Hopkins Medicine Baltimore United States

1 Stony Brook University* Stony Brook United States

1 Cornell University Ithaca United States

1 Zero to Three Washington United States

1 Mississippi State University Starkville United States

1 Creative Commons Mountain View United States

1 Otsuka (United States)* Princeton United States

21
 1 University of Notre Dame Notre Dame United States

1 University of California, San Diego San Diego United States

1 State University of New York Albany United States

1 Northeastern University Boston United States

1 The University of Texas at Austin Austin United States

1 University of California‒Los Angeles Los Angeles United States

1 Loyola University Medical Center Maywood United States

1 North Carolina State University Raleigh United States

1 Michigan State University East Lansing United States

1 Graduate School USA Washington United States

1 Intel (United States) Santa Clara United States

1 University of California, Davis Davis United States

1 University of Michigan–Ann Arbor Ann Arbor United States

1 Block Engineering (United States) Southborough United States

1 Capital University Bexley United States

1 New York Institute of Technology New York United States

1 Case Western Reserve University* Cleveland United States

1 Purdue University West Lafayette* West Lafayette United States

1 University of California, Berkeley Berkeley United States

1 The University of Texas MD Anderson Cancer Center Houston United States

1 Johns Hopkins University Baltimore United States

1 Carnegie Mellon University Pittsburgh United States

1 Amgen (United States) Thousand Oaks United States

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 22

1 Stanford Medicine Stanford United States

1 Rensselaer Polytechnic Institute Troy United States

1 Boston Children’s Hospital* Boston United States

1 Center for Information Technology* Bethesda United States

1 University of California, San Francisco* San Francisco United States

1 Pfizer (United States) New York United States

1 King University* Bristol United States

1 ORCID Bethesda United States

1 FuelCell Energy (United States) Danbury United States

1 Lamar University Beaumont United States

1 Applied Materials (United States) Santa Clara United States

1 University of Chicago Chicago United States

1 Rutgers, The State University of New Jersey New Brunswick United States

1 University of Memphis Memphis United States

1 University of North Carolina at Chapel Hill Chapel Hill United States

1 Southern California University for Professional Studies Irvine United States

1 Optica Washington United States

1 The Ohio State University Wexner Medical Center Columbus United States

1 University at Buffalo, State University of New York Buffalo United States

1 Google (United States)* Mountain View United States

1 University of Arizona Tucson United States

1 Unchained Labs (United States) Pleasanton United States

1 Boston College* Boston United States

23
 1 Health First* Rockledge United States

1 University of Akron Akron United States

1 Hunter College New York United States

1
DeepSeek’s announcement roiled US markets, leading to a 3 percent decline in the NASDAQ
composite and a 17 percent drop in NVIDIA shares, erasing $600 billion in value. It was the largest
single-day loss of a company in US history—a figure equivalent to 65 percent of the annual US defense
budget. For more information: https://www.cnn.com/2025/01/27/tech/deepseek-stocks-ai-
china/index.html.
2
See Appendix A for details on each of the five DeepSeek papers.
3
Note: For the first four papers, author lists were consistent between PDF and arXiv metadata. However,
for the fifth paper (arXiv:2501.12948), we found discrepancies between authors listed in the PDF and the
arXiv metadata, so we included all unique authors from both sources to ensure comprehensive coverage.
Additionally, 11 authors across all papers could not be found with OpenAlex profiles and were excluded
from the analysis.
4
OpenAlex is a tool hosted by OurResearch, a nonprofit focused on open science tool development.
5
See Appendix B for a full list of names in the Key Team.
6
The h-index captures the number of publications with at least h citations (i.e., an h-index of 13 implies
13 papers cited at least 13 times), while the i10-index counts how many works have at least 10
citations—useful for gauging consistency across a body of work.
7
See Appendix B for the full dataset.
8
While 211 authors were included in the full bibliometric analysis, the affiliation-based breakdowns in
this chart total 201 due to 10 individuals with no available institutional data in OpenAlex. These 10
authors also had very limited bibliometric profiles, with an average of just 4.4 publications, 8.6 citations,
and near-zero recent citation activity—suggesting that they are likely junior researchers or early-career
contributors. Their omission from affiliation analysis does not significantly affect aggregate findings but is
noted here for transparency.
9
For more information about institutional relationships in OpenAlex, see https://docs.openalex.org/api-
entities/institutions/institution-object.

Amy Zegart and Emerson Johnston | A Deep Peek into DeepSeek 24