An AI Alchemist and His DeepSeek Journey

How a hedge fund manager, Wenfeng Liang took self-funded DeepSeek to the world stage

An Unexpected Homecoming

The lunar new year celebration in Miling Village, Zhanjiang, Guangdong (a village in southern China), has become an unexpected social media sensation. The village, usually quiet, now bursts with red festive banners reading, "Warmly welcome Liang Wenfeng home." The last Zhanjiang native to receive such a reception was Hongchan Quan, the then 14-year-old gold medalist who stunned the world with the highest-ever score in women’s 10-meter diving at the Tokyo Olympics.

Wengfeng Liang’s story is just as remarkable. A quiet and studious guy with glasses, born in 1985 (the same year as Sam Altman, founder and CEO of OpenAI), Liang made waves in the global tech community with his self-funded open-source AI large language model (LLM), DeepSeek.

In a matter of weeks, DeepSeek has gained worldwide attention. In late December, the DeepSeek-V3 open-source model was released, integrating three groundbreaking technologies: FP8, MLA (multi-head potential attention), and MoE (mixture of expert) architecture, significantly boosting performance and efficiency. Wall Street took notice, as each subsequent release surpassed the previous one. On January 20, 2025, DeepSeek introduced DeepSeek-R1, a model designed for tasks like mathematics, coding, and logic, achieving performance levels comparable to OpenAI’s GPT models. Just a week later, on January 27, DeepSeek unveiled the Janus Pro 7B and 1.5B models, notable for their ability to run on consumer-grade hardware like AIPC, credits to their smaller parameter sizes.

What’s even more surprising is the innovative approach to training DeepSeek’s models. Although the company did not disclose training costs for DeepSeek-R1, reports indicated that it was trained using a specialized version of Nvidia check that is available only in China due to U.S. export controls. This led to significantly lower training costs compared to other LLMs. While industry experts have different opinions and estimates on the training cost, most agree with the innovative approach DeepSeek team took with limited resources available.

The Unforeseen Rise

The ripple effects of DeepSeek have shaken the tech world, even leading investors to question NVIDIA’s growth strategy, which contributed to a sharp one-day drop in Nvidia’s market cap.

DeepSeek has paved the way for more cost-effective training methods, and lowered the barrier to experimenting and building AI. DeepSeek has quickly made a splash and is being adopted across the industry. For instance, small-scale experiments at institutions like the University of California, Berkeley, and the Hong Kong University of Science and Technology have successfully validated DeepSeek’s potential. These experiments show that smaller, more accessible models can also benefit from DeepSeek’s innovation. This is sure to spark greater interest among small research labs, startups, and smaller institutions eager to get involved in AI development.

What’s more, the leading AI companies also started announcing support for R1, including Microsoft Azure, Amazon Bedrock, Nvidia NIM, Perplexity, Hugging Face, just to name a few. As more companies join in, the pace of innovation and real-world applications will only accelerate.

The meteoric rise of DeepSeek has made Liang and his team uneasy according to insiders. They wish for the public’s attention to cool down so they can return to focusing on research and development. DeepSeek, established in 2023, has remained relatively low-profile, even flying under the radar of venture capitalists. Unlike many other AI companies, DeepSeek has been entirely self-funded by Liang and has never sought outside investment.

From Quant Trading to Open-Source AI: An Unconventional Path

Before founding DeepSeek, Liang and his team had already invested in more than 10,000 GPUs for model training for quantitative trading. Tracing back to DeepSeek’s origins, you inevitably connect it to the quant fund, Magic Square Quantitative, which Liang founded with another college friend. DeepSeek was essentially incubated within Magic Square Quantitative, and its AI roots were ingrained in the fund's core strategy.

According to the official website of Magic Square Quantitative, AI is the core strategy of the fund. The fund proudly claims to be a leader in AI-driven quantitative trading.

“Since 2008, we’ve utilized machine learning and other technologies to explore fully automated quantitative trading. On October 21, 2016, we launched our first trading position powered by deep learning. Over the following year, Magic Square Quantitative became fully AI-based, establishing itself as a leader in AI trading not only in China but also globally.”

Magic Square’s AI training platform, Firefly, has also gained a strong reputation. Firefly No. 2, introduced a unique hardware and software architecture that doubled computing power and sped up models by 50-100%. The platform has been crucial in supporting strategy research and model testing without being limited by computing power.

Liang believes that fund managers are essentially programmers and servers. Though he rarely gives public interviews or speeches, in a keynote titled "The Future of Quantitative Investment in China from the Perspective of a Programmer", Liang explained,

"Quant companies don’t really have fund managers—fund managers are just a bunch of servers. While investment decisions made by people are based on feelings, decisions made by computer programs are based on science, always leading to the best solution."

Magic Square Quantitative stands out in the industry because of its sophisticated mastery of AI-driven productivity. By 2019, the company’s assets surpassed 10 billion RMB (~1.39B USD), and by 2021, it had become the first Chinese quantitative private equity (quant PE) firm to break through the 100 billion RMB (~13.9B USD) mark. It quickly gained recognition as one of the "Four Kings" of quant PE in China.The success from Magic Square enabled Liang to make the decision not to raise outside funds when founding DeepSeek.

Born in 1985, Liang made his first 10 billion RMB in his 30s through his expertises in technology. Unlike other AI model companies that rely on hefty rounds of funding, Liang’s ability to finance DeepSeek independently allowed the company to focus solely on technology development. The laser focus on technology is also why Liang chose the open-source model for DeepSeek. It’s an intriguing contradiction: a highly successful hedge fund manager with top-tier earning potential choosing to invest heavily in research and model capabilities rather than focusing on commercial KPIs.

Why the open-source approach? Liang explained

“The AI industry is still in its early stages, and closed-source models are hard to commercialize in the short term. Creating an open-source ecosystem, free from profit-driven pressures, will encourage more developers to participate, fostering an AI ecosystem that benefits wider audience and accelerates technological progress.”

Liang’s love for mathematics and scientific research is reflected in the names of his two companies. Magic Square Quantitative, which was the foundation of his AI ventures, is a traditional Chinese game involving number arrangement, where the sum of numbers in each row, column, and diagonal is equal. It’s an apt metaphor for the precision and analytical focus that underpins both his work and products. Similarly, DeepSeek—whose name reflects a commitment to deep, meaningful exploration—aims to develop artificial general intelligence (AGI) with broad cognitive capabilities, constantly pushing the boundaries of what AI can achieve.

The technical roots of Magic Square and DeepSeek are deeply intertwined, particularly in areas like algorithm development, big data processing, and high-performance computing. DeepSeek builds on the AI infrastructure developed by Magic Square Quantitative, but it operates as an entirely independent entity focused on advancing AI technology.

“The moat formed by closed-source models is short-lived”

Liang's geek spirit is evident in all aspects of his work, including his approach to AI development. A former employee at DeepSeek once described him as "always looking unkempt" but deeply passionate about his work. Despite his self-made wealth, Liang immerses himself in the technical details of his projects, from adjusting learning rates to defining network parameters. His commitment to research is evident in every part of the company, where even the smallest details are personally overseen by him.

It’s also worth noting that Liang has never studied abroad. A graduate of Zhejiang University’s Department of Electronic Engineering, he represents a departure from many other Chinese fund managers, who often studied and worked overseas.

The DeepSeek team is also high caliber. According to the website of Magic Square Quantitative, the company recruits top-tier talent in mathematics, physics, informatics, and AI. Its strategy and development team includes gold and silver medalists from the International Olympiad in Mathematics and Physics, as well as AI leaders and PhDs in fields like Topology, Statistics, and Operations Research. Collaboration across disciplines is central to their work, allowing them to solve complex challenges in deep learning, big data modeling, and parallel computing.

A now-deleted article profiled the DeepSeek team circulated online in China. The article revealed that most of their R&D team consists of PhDs from top Chinese universities like Tsinghua, Peking University, and Fudan University—and most of them have never studied abroad. One former employee shared that the atmosphere at DeepSeek is research-focused, with a flat organizational structure and a youthful, curious team passionate about AI development.

In an interview, Liang shared his belief that the true value of a company lies in its team, especially when facing disruptive technology. “The moat formed by closed-source models is short-lived,” he said, emphasizing that DeepSeek’s success is rooted in its innovative culture and collaborative team. On the official website of Magic Square Quantitative, one can find his personal motto

"AI expands the boundaries of our capabilities and inspires our imagination and creativity."

His techno idealism, uninfluenced by capital nor politics, stems from a deep curiosity and belief in the power of open innovation—qualities that have propelled DeepSeek into the global spotlight. While other AI companies grapple with balancing capital and performance, Liang’s story stands out - one that demonstrates his belief and commitment that open collaboration is crucial for the long term health of the AI ecosystem.

The AI “Alchemist”, A Story of Passion and Belief

In China, training LLM is often referred to as "alchemy," a term borrowed from ancient legends where immortals or martial arts masters used mystical ingredients to create magical elixirs. These potions were believed to grant immortality or enhance martial prowess, but the process was full of uncertainty and difficulty, making the final product incredibly rare and valuable. This mirrors the process of training LLMs: under the influence of data, engineering, training methods and resources, the effectiveness of each model is uncertain until it’s "perfected”. The journey is arduous, and only after much trial and error does the model become truly “refined”. This is where Liang and DeepSeek's story resonates. It's the tale of a young, successful hedge fund manager, driven by a passion for technology and innovation, who has propelled his AI company onto the global stage. This time, with a higher purpose, a relentless pursuit of open-source AI advancement.