Talent & Culture Strategy at DeepSeek: High Density, Flat Hierarchy, No KPIs

Ross Geller

·April 15, 2026

This article is part of MokaHR's Talent & Culture Strategy series, which profiles how leading companies build their people strategies.

Introduction

On 20 January 2025, a Chinese research lab most Western technology investors had never heard of released an open-source reasoning model called R1. Within seven days, DeepSeek's app had surpassed ChatGPT to become the number one free download on the United States iOS App Store. US technology stocks lost more than $1 trillion in market capitalisation amid what analysts called a "Sputnik moment" for the AI industry. Nvidia alone lost nearly $600 billion in market value in a single day.

The model had been trained using 2,048 Nvidia H800 GPUs at a reported cost of $5.6 million — orders of magnitude cheaper than the frontier models produced by OpenAI, Anthropic, and Google DeepMind. DeepSeek was not a state-backed mega-project. It was not a spin-out from a US research lab. It was a 150-person team in Hangzhou, funded by a Chinese quantitative hedge fund, staffed almost entirely by fresh graduates from Chinese universities, operating with no formal KPIs and no management hierarchy. Its founder, Liang Wenfeng, rarely gave interviews, held 1% of the equity, and had been described by his own colleagues as "not at all like a boss and much more like a geek".

For HR leaders studying the AI industry, DeepSeek is the single most important data point of the last two years — not because its model was better than the frontier Western labs, but because the combination of its organisational design, hiring philosophy, and capital structure delivered competitive frontier output at a fraction of the headcount and spend. The company is the strongest modern evidence that elite technical work can come from very small, very flat, very young teams — if the hiring filter is right.

The question worth studying is what specifically DeepSeek does that allows 150 people to produce research outputs competitive with organisations ten to fifty times their size.

Detail	Data
Founded	May 2023, Hangzhou
Headquarters	Hangzhou, China (offices in Beijing)
Employees	~150–200 at time of R1 release (January 2025), expanding
Parent funding	High-Flyer Quant (~$8 billion AUM) — no external VC
Core business	Frontier AI research and open-source model releases
Founder's equity stake	~1% (99% held by High-Flyer partnership)
DeepSeek-V3 training cost	~$5.6 million on 2,048 Nvidia H800 GPUs
Core technical innovation	Multi-head Latent Attention (MLA), Mixture-of-Experts
Team composition	Predominantly fresh graduates from top Chinese universities

How does DeepSeek attract and hire talent?

DeepSeek's hiring philosophy is the most extreme example of "high talent density, low conventional credentials" of any major AI lab. Liang Wenfeng has been explicit about this in his rare public interviews: the company's hiring standard is "passion and curiosity", not prior experience, and core technical roles are filled primarily by fresh graduates or those one to two years into their careers.

The inversion of conventional AI hiring

Almost every frontier lab in the United States — OpenAI, Anthropic, Google DeepMind, Meta — filters heavily on publication count, PhD credentials, or prior industry track record. DeepSeek deliberately inverts this. As Liang told 36Kr in 2024, "DeepSeek has not hired anyone particularly special and employees tend to be locally educated", and the company sought employees "who had ability and passion rather than experience". Core authors of DeepSeek's technical papers are drawn heavily from Tsinghua, Peking University, Zhejiang University (Liang's alma mater), and the Chinese Academy of Sciences ecosystem.

The bet is that curiosity and learning velocity, at the scale of a 150-person team, produce more original research than experience-heavy hiring at the scale of a 3,000-person lab. A 2025 Stanford and Hoover Institution joint analysis found that despite the youth of the team, DeepSeek authors had accumulated substantial academic output: an average of 61 published papers and 1,059 citations per researcher, with the 31 most core authors averaging 70 papers and 1,554 citations each. The team is young, in other words, but not untrained.

Autonomy as the recruiting proposition

The most distinctive element of DeepSeek's recruiting pitch is not compensation but autonomy. Liang has described the working environment as one where "anyone on the team can access GPUs or people at any time. If someone has an idea, they can access the training cluster cards anytime without approval." The company has no fixed hierarchies and no separate departments — people collaborate across teams "as long as there's mutual interest". For young researchers at the beginning of their careers, the ability to run large-scale compute-intensive experiments without bureaucratic approval is a recruiting asset that no established lab can easily match.

Liang has pointed to Multi-head Latent Attention (MLA), a core architectural innovation behind DeepSeek's training-cost advantages, as a direct product of this structure. MLA originated as a young researcher's personal interest; the company formed a team around the idea, committed months of compute, and produced an architectural change that dropped training costs substantially. In a conventional lab, the same proposal from a junior researcher would typically require multiple layers of approval before any compute allocation.

This approach contrasts sharply with OpenAI's reactive retention model, which uses $1.5 million bonuses and accelerated vesting to retain senior staff against Meta's nine-figure offers. DeepSeek's talent proposition is structurally different: pay competitive packages (the top AGI researchers earn up to 1.54 million yuan annually on a 14-month pay structure), but win on autonomy, problem access, and intellectual ambition rather than financial defence.

Compensation that is competitive by Chinese standards

DeepSeek's compensation is well above the Chinese AI industry average but modest by Silicon Valley standards. According to February 2025 reporting in China Fund News, AGI deep-learning researchers earn between 80,000 and 110,000 yuan per month on a 14-month structure, for annual packages of up to roughly 1.54 million yuan (approximately $213,000 at then-current exchange rates). Client-side engineers — the lowest-tier technical roles — start at 200,000 to 300,000 yuan annually. For comparison, the average AI engineer in China earned approximately 380,000 yuan per year according to Glassdoor data from the same period.

The packages are competitive for attracting top Chinese graduates who might otherwise pursue opportunities at Alibaba, ByteDance, or Silicon Valley firms, but they are not the primary recruiting lever. Multiple accounts have described young researchers choosing DeepSeek over Nvidia and other US firms primarily because of the combination of research autonomy, proximity to family, and the opportunity to take on significant roles early in their careers — the same combination that pulled Zizheng Pan back from an Nvidia internship in 2023.

The open-source signal

DeepSeek's commitment to open-sourcing all its models functions as a recruiting signal as well as a competitive strategy. Liang has argued that "in the realm of disruptive technologies, closed-source methods serve only to delay progress temporarily" and that "the technology should not be controlled by only a few people and companies". For researchers early in their careers, open-source commitment means their work will be publicly visible, citable, and buildable-upon — a meaningful career consideration that compensates for the lower global brand recognition of a Chinese AI lab relative to an American one.

📄 2025 AI Recruitment Casebook DeepSeek's ability to produce frontier research with a 150-person team challenges every assumption about headcount being a prerequisite for AI capability. To see how companies across 10 industries are rethinking talent density, hiring filters, and organisational structure in the AI era, download MokaHR's full AI Recruitment Report. Download the free report →

Trusted by: Tesla · NVIDIA · McDonald's · Nestlé · Schneider Electric

How does DeepSeek develop and manage employee performance?

Performance management at DeepSeek is structured around a deliberate absence: no formal KPIs, no performance-review cycles in the conventional sense, no ranking, and no "horse-racing" internal competition of the kind that defines ByteDance and Tencent. The model is unusual not only for a Chinese technology company but for any frontier AI lab globally.

"Natural division of labour"

Liang has described DeepSeek's operating model as one of "natural division of labour" — research groups form around specific goals, each member works on the part of the project they are best at, and difficulties are discussed collectively. There are no fixed hierarchies, no rigid roles, and no mandatory team boundaries. Collaboration across groups happens when people identify mutual interest, not when managers approve it.

This structure is most comparable to Anthropic's approach to integrated research and engineering, where the lines between disciplines are deliberately blurred. The difference is scale: Anthropic has roughly 2,500 employees; DeepSeek has around 150–200. At that scale, the absence of formal structure is feasible in a way it would not be in a larger organisation.

No KPIs, no commercial pressure

DeepSeek's funding structure is the critical enabler of its performance model. Because the company is fully funded by High-Flyer, Liang's quant hedge fund parent, there is no external investor pressure to generate near-term returns. Liang holds only 1% of the equity — the other 99% is held by the Ningbo High-Flyer Quantitative Investment Management Partnership. The company has no fundraising plans and no public commitments to commercial milestones. This gives DeepSeek the unusual freedom to, in Liang's words, treat "basic research" as "a low-return business, requiring long-term dedication" rather than a commercial product pipeline.

The obvious parallel is to early OpenAI, when Sam Altman famously told investors: "We have no idea how we may one day generate revenue. We have made a soft promise to investors that, 'Once we've built a generally intelligent system, basically we will ask it to figure out a way to generate an investment return for you.'" Both companies shared small teams (OpenAI had roughly 150 employees in 2019), unconventional governance structures, and research-over-commerce prioritisation. The difference is that DeepSeek's funding arrangement is durable — it is not dependent on an external investor base — while OpenAI's shifted rapidly once ChatGPT created commercial pressure.

Development through problem access

Because DeepSeek has no formal performance-review process, development happens through exposure to increasingly difficult problems rather than promotion ladders. The 150-person team structure means that even junior researchers work directly alongside core authors of the company's technical papers. This is fundamentally different from the training-and-development programmes run at Google DeepMind, where a 45-day interview process and formal research mentorship structures exist precisely because the scale of the organisation requires them. DeepSeek's development model is closer to a traditional academic lab: you learn by working on real problems alongside people who are solving them.

The profitability discipline

A subtle but important detail: despite triggering an AI price war in China with its DeepSeek-V2 release in May 2024, the company has remained profitable since 2024 — in contrast to every frontier US lab, which operates at substantial losses. This is partly the product of the efficiency-first architectural choices (MLA, Mixture-of-Experts) and partly the product of being funded from High-Flyer's existing R&D budget rather than venture capital. For employees, the signal is that the company is durable without needing to make aggressive commercialisation bets — a retention asset in a market where many peers are visibly straining to justify their valuations.

What can HR leaders learn from DeepSeek's approach?

Most Western organisations cannot replicate DeepSeek's funding structure, and many cannot replicate its talent density in frontier research. But the operational choices that produced a $5.6 million frontier model with a 150-person team are more portable than they initially appear.

Hire for curiosity at the start of careers, not for track record. DeepSeek's decision to recruit primarily fresh graduates and researchers with one to two years of experience is a high-variance bet, but at the scale of a 150-person team, high variance is an asset. The hires who work out produce disproportionate breakthroughs (MLA is the clearest example); the hires who do not are relatively cheap to correct. The generalisable lesson: in functions where original work matters more than incremental delivery, filtering for trajectory and curiosity rather than credential produces better outcomes than experience-heavy hiring. AI-powered screening tools can help HR teams operationalise this by identifying signals of learning velocity and intellectual engagement at the application stage, rather than relying on proxy credentials.

Reduce approval friction on resources that limit breakthrough work. DeepSeek's most distinctive policy is that any team member can access GPU compute without approval. The principle generalises: in every knowledge-work organisation, there is some critical resource — compute, data access, legal sign-off, budget — whose approval friction determines how much experimentation actually happens. Reducing approval gates on that specific resource, for trusted staff, is often the highest-leverage operational change an organisation can make. Most companies over-govern their scarce resources; DeepSeek's decision to under-govern them for a small, high-trust team has been one of the biggest drivers of its output.

Keep teams small enough that structure is unnecessary. The 150-person threshold matters. Above it, informal coordination breaks down and organisations start adding hierarchy, KPIs, and departmental boundaries to compensate. DeepSeek has deliberately resisted this transition for as long as possible. For HR leaders running high-performing teams anywhere, the lesson is that the transition from small-team informality to large-team structure should be delayed as long as it remains feasible, because the productivity tax of adding structure is almost always larger than the marginal benefit. Workforce analytics platforms can help HR teams identify the point at which informal coordination is genuinely breaking down, rather than adding structure reactively.

Decouple hiring from commercial pressure where funding allows. DeepSeek's independence from venture capital is unusual and unreplicable for most companies, but the principle — separating the hiring decision from near-term revenue pressure — can be approximated. Organisations with strong balance sheets can fund research, platform, or R&D teams on multi-year horizons rather than quarterly budgets, which produces dramatically different hiring filters and retention outcomes. This contrasts with OpenAI's hiring expansion to 8,000 employees driven by enterprise revenue pressure, which is optimised for shorter-horizon commercial outcomes.

What is it like to work at DeepSeek?

Working at DeepSeek in 2025 and 2026 means working at one of the smallest frontier AI labs in the world, with a level of research autonomy that staff at larger labs have described as enviable. The working style is closer to a well-funded academic group than a conventional technology company: flat, collaborative, focused on papers and open-source releases, with the CEO described as spending most days "reading papers, writing code, and participating in group discussions" rather than running meetings.

The workforce is young. Most staff are in their twenties, drawn from Tsinghua, Peking University, Zhejiang University, and the Chinese Academy of Sciences. Many are fresh graduates or interns; some joined instead of accepting offers from Nvidia, Google, or Microsoft. The team has been described as "rank-less and extremely flat", with members divided into research groups according to specific goals rather than permanent team assignments. Compensation is competitive by Chinese AI industry standards but not by Silicon Valley standards — the draw is the work, not the pay.

The company has also been the focus of enormous public and political attention since the R1 release. Following the model's January 2025 launch, thousands of young Chinese jobseekers applied to DeepSeek, with some driving across the country to present themselves in person at the Hangzhou headquarters. The Communist Party secretary of Guangdong province publicly commended the company. That visibility has been an asset for recruitment but has also raised the pressure on the organisation to maintain its original operating model as it scales.

The challenges are also real and acknowledged. As DeepSeek expanded from 150 employees to more than 200 through 2025, the informal coordination model that worked at 150 will eventually require adaptation. The company's competitive intelligence is also now a geopolitical concern: US chip export controls have forced DeepSeek to accelerate innovation around compute efficiency, but also limit the compute ceiling available to the team. Liang has acknowledged that "all strategies are products of the past generation and may not hold true in the future" — an unusually honest admission for a founder at the peak of industry attention.

One particularly interesting observation: DeepSeek has been less vulnerable to external poaching than comparable US labs. Partly this is because the company's success has not been tethered to a small number of "legendary" scientists — architectural breakthroughs like MLA are credited to young researchers by name, distributing the intellectual ownership across the team. Partly it is because Chinese AI talent that chose to return from the US is by definition mission-aligned with building Chinese frontier AI. And partly it is because DeepSeek's compensation packages, while modest by Meta standards, are well above the Chinese industry average and funded durably by High-Flyer.

Whether the DeepSeek model survives the transition from 200 to 500 employees, from single-mission research to multi-product deployment, and from its current funding structure to whatever follows is an open question. But the evidence from 2024–2026 is that a small, dense, autonomy-rich team with the right hiring filter can compete with organisations ten to fifty times its size — and that is a finding every HR leader in the technology sector should take seriously.

Frequently asked questions

How many employees does DeepSeek have? DeepSeek had approximately 150 to 200 employees at the time of its January 2025 R1 release, and had expanded to fill more than 50 open roles across Hangzhou and Beijing by February 2025. The company operates with a deliberately small, dense team drawn primarily from top Chinese universities, funded by its parent quant hedge fund High-Flyer rather than external investors.

What is DeepSeek's hiring philosophy? Founder and CEO Liang Wenfeng has stated that DeepSeek's hiring standard is "passion and curiosity", with ability and potential valued above prior experience. Core technical roles are filled primarily by fresh graduates or researchers with one to two years of experience, drawn from universities including Tsinghua, Peking University, and Zhejiang University. A 2025 Stanford and Hoover Institution analysis found that despite the youth of the team, DeepSeek authors had accumulated an average of 61 papers and 1,059 citations per researcher.

Does DeepSeek use KPIs? No. DeepSeek operates without formal KPIs, without rigid hierarchies, and with a "rank-less and extremely flat" organisational structure. Research groups form around specific goals, and any team member can access GPU compute resources without approval. Liang Wenfeng has publicly attributed key architectural innovations, including Multi-head Latent Attention, to this bottom-up structure.

What does DeepSeek pay its engineers? According to China Fund News reporting from February 2025, DeepSeek offers 14-month pay packages with AGI deep-learning researchers earning between 80,000 and 110,000 yuan per month — equivalent to an annual income of up to 1.54 million yuan (approximately $213,000). The lowest listed technical role, client-side engineer, starts at 200,000 to 300,000 yuan per year. These packages are substantially above the Chinese AI industry average of 380,000 yuan annually, though well below comparable Silicon Valley frontier-lab compensation.