DeepSeek-V3 won three out of six LLM large language model tests, particularly in reasoning tasks such as the MATH 500 math test (90.2%) and the Codeforces and SWE programming tests.
Surprisingly, while DeepSeek-V3’s performance is comparable to the top-end large-scale model GPT-4o (closed source), it only cost $5.58 million to develop and less than 1/20th the training cost. It took Google and OpenAI several years, billions of dollars, and tens of thousands of cutting-edge GPUs to achieve the same result.
The launch of DeepSeek-V3 strangely coincided with the emergence of the sixth-generation aircraft in the Chinese Air Force, which the Americans compared to the USSR’s launch of the first artificial satellite. However, what really shocked Silicon Valley was not the high performance and low cost of DeepSeek-V3, but the innovative spirit of the model. Until then, it was believed that technological breakthroughs were characteristic of the United States, while China was only good at applications. Even more surprising was that DeepSeek-V3 was not backed by Asian tech giants like Tencent or Ali, but by a modest private equity fund, Magic Square Quantitative, which was permanently headed by tech geek Liang Wenfeng.
Company secret
DeepSeek is a company shrouded in mystery. Its head office in Beijing occupies an entire floor, 2,100 square meters. The address is known, but the security guards downstairs won’t even tell you the floor number, and there’s no sign with the name in the lobby. Moreover, additional security has recently been posted on the floor. More than a hundred people try to get into DeepSeek every day, but, as a rule, they are unsuccessful. The same thing happens with phones and emails: everyone knows them, but it’s impossible to contact the employees. Although until recently there was no excitement.
“I was offered a job at this company in 2023, but I didn’t pay any attention to it,” says a leading Chinese AI developer . “I would be happy to move there now, but who would offer it?”
Magic square
Liang Wenfeng graduated from Zhejiang University in Hangzhou in 2008, where he studied artificial intelligence. However, after graduation, the young specialist did not go to work for a large IT company, but went to Chengdu, where he became interested in the ideas of mathematician Jim Simons about algorithmic trading. This approach uses sophisticated mathematics to analyze transactions, which allows you to develop better algorithms and models for making decisions about buying or selling stocks and other financial instruments.
This is where the so-called quants work – analysts who use methods taken from signal processing, game theory, the Kelly criterion of gambling, market microstructure, econometrics and time series analysis. These are the best mathematicians in the world – usually former participants of international Olympiads. They say that many theorists working in string theory have left for Jim Simons himself, the founder of Renaissance Technologies – the number one company in the field of algorithmic trading. This is the highest-paid job, with salaries there amounting to millions of dollars.
The main companies developing this area were located in the US and UK, but in China it was a new market. After two years of hard work and many failures, Wenfeng’s algorithms started working, which brought him about $70 million. And in 2015, Liang, together with classmates from Zhejiang University, founded the company Magic Square Quantification.
The idea of the young people was to use their knowledge of artificial intelligence to create algorithms that would allow them to become a world-class company, a kind of Chinese Renaissance Technologies. Just a year later, the startup launched the first real AI-based trading, and then transferred all trading strategies to this technology. At its peak, high-frequency arbitrage using AI brought in 300% profit! In 2021, Magic Square Quantification’s assets under management reached 1 trillion yuan ($137 billion), making the company a Chinese Quant giant – Liang Wenfeng’s dream come true.
From QUANT to AGI
AI trading strategies require enormous computing power. In particular, as the model parameters increase, the demand for GPU computing power increases. It is logical that Liang’s next step was computing power.
In 2019, Magic Square invested 200 million yuan ($28 million) to build the Firefly One AI computing cluster, equipped with 1,100 GPUs, just as Musk was coming up with the concept for his Tesla Dojo cluster. A few months later, when Nvidia released its latest A100 chip, Wenfeng became the first person in the Asia-Pacific region to get the card.
In 2021, he invested another 1 billion yuan ($137) to build Firefly II with 10,000 A100 cards, which had the computing power of 760,000 personal computers. The cluster was larger than 10 basketball courts, and Liang became the largest private buyer of Nvidia A100 chips in Asia. The era of large LLM language models that require such power had not yet arrived, and many thought the founder of Magic Square Quantification was crazy.
DeepSeek
But Liang Wenfeng is definitely not crazy — this man with phenomenal mathematical abilities thinks many moves ahead. And talks very little. We know for sure that in July 2023, he founded a company in Hangzhou called DeepSeek, the main goal of which is to develop large LLM language models, and ideally AGI (Artificial General Intelligence) — a hypothetical form of AI that has the ability to learn and solve problems similar to human intelligence. For what purpose?
The first version is idealistic. Liang wants to make the technology accessible and give it to his country and the rest of the world, thereby turning China from imitators to creators.
“For 30 years we lived by the logic of ‘the West invents – China copies’. But in the era of AI, this is a road to nowhere. If we do not develop original technologies, we will remain eternal outsiders, as in the case of chips,” is one of the few direct quotes from Liang.
The second version is practical. The previous algorithms stopped working. Magic Square’s assets under management have fallen from 1 trillion yuan in 2021 to 20 billion in 2025. DeepSeek’s breakthrough in generative AI could revolutionize algorithmic trading. And Liang sees AGI as the key to a new level of financial strategies. The truth, as usual, is somewhere in the middle.
Dream job
DeepSeek employs about 160 people, split between two offices, Beijing and Hangzhou. Even if they were all in Beijing, there would be at least 9 square meters of space per person. Such a huge office is rare for a startup. In fact, it is even more spacious. When asked what they do in different cities, DeepSeek answers in Chinese: “Beijing is blooming, and Hangzhou is bearing fruit.” To add to the confusion, the core employees of DeepSeek and Magic Square Quantitative are the same people.
The startup displays a typically Asian mix of hardline authoritarianism and absolute democracy. Liang controls 84.29% of the shares and has a decisive say, but the company is also extremely relaxed. Wenfeng prefers to hire former students with master’s degrees and experience in a large language model, mainly from Peking University and Tsinghua University.
90% of the team are graduates of Chinese universities who have not had any internship abroad. “We cultivate geniuses, not poach them,” says Liang. “The world’s 50 greatest talents may not live in China, but we can cultivate them ourselves” is Wenfeng’s second famous statement on the same topic . Interestingly, the employees have no KPIs, no one marks their arrival and departure times, and they set goals based on their own interests.
In terms of its personnel structure, DeepSeek has broken all traditional norms by hiring mostly young people (the key technical positions are filled by people who have only been out of university for a year or two) and valuing their ability and enthusiasm over their qualifications and experience. Liang believes that students and graduates are at the peak of their abilities and have fewer external constraints, making them better suited to achieve breakthrough results. Wenfeng is quite generous: his company’s salaries are twice as high as those of its competitors. For example, a hired deep learning researcher or research engineer is immediately paid about 1.54 million yuan ($12,000) per year — young professionals are unlikely to be offered more than 800,000 elsewhere.
Liang Wenfeng himself is less like a boss and more like a geek: he comes up with algorithms, writes code, and is involved in recruiting staff. Any employee can contact him directly and discuss code and algorithms. The company does not have a hierarchical management model: if an interesting idea arises during the research process, its author can involve people in analysis and freely use extensive computing resources for implementation.
“At DeepSeek, everything is done from the bottom up,” Liang explains the company’s policy . “We usually don’t divide functions up front, but rather apply a natural division of labor. Each employee has their own unique growth experience, everyone has their own ideas, and we don’t need to pressure people. If someone encounters a problem during the research process, they invite their colleagues to discuss it. However, when an idea shows potential, we also allocate resources from the top down.”
Pure art
From the start, DeepSeek has stated that its core mission is to explore the nature of artificial general intelligence. Few in China’s AI industry have dared to set such wild goals. As a result, in the past few years, while many major AI makers have been working to attract users and make money, Wenfeng has been pursuing fundamental research that so far appears unprofitable. According to the scientist, “innovation is not entirely dependent on business; it also requires curiosity and creativity.” And while Chinese companies were previously bound by commercial inertia, DeepSeek has freed itself from this constraint.
The startup is financed from Magic Square’s revenues: 1.5% of the total investment portfolio under the company’s management and 20-25% from profits from exceeding target indicators. Liang says he does not need external financing, and no wonder: according to Bloomberg Research, DeepSeek is worth up to $330 billion. This price is almost equal to the total current valuation of China’s AI Six Dragons: Baidu, Alibaba, Tencent, Huawei, Xiaomi and ByteDance, and therefore there is no end to those wishing to invest in the company.
For example, Zhu Xiaohu, managing partner of Jinshajiang Venture Capital, said that Liang Wenfeng has changed his skepticism about artificial general intelligence (AGI) and if DeepSeek opens funding, he will definitely invest in this startup because it is very important to “witness the birth of human AGI.” “The price does not matter much now, the main thing is to participate in this incredible event,” he said.
“I couldn’t find a business reason to start DeepSeek even if you asked me,” Wenfeng explains . “Because it doesn’t make business sense. Basic science research has a very low return on investment. When OpenAI’s early investors put money into it, they certainly weren’t thinking about how much they would make. It was more that they really wanted to do it.”