A Heartwarming Homecoming
The lunar new year celebration in Miling Village, Zhanjiang, Guangdong—a serene village in southern China—has unexpectedly captured social media attention. Festive red banners proclaiming “Warmly welcome Liang Wenfeng home” adorn the village, reminiscent of the reception given to Hongchan Quan, the young gold medalist who astonished the world with her diving skills at the Tokyo Olympics.
Wenfeng Liang’s narrative is equally remarkable. Born in 1985, the same year as OpenAI’s founder Sam Altman, Liang emerged as a significant figure in the global tech landscape with his self-funded open-source AI large language model (LLM), DeepSeek.
In just weeks, DeepSeek garnered international acclaim. The release of the DeepSeek-V3 open-source model in late December introduced three pioneering technologies: FP8, MLA (multi-head potential attention), and MoE (mixture of experts) architecture, enhancing both performance and efficiency. Wall Street took note as each release surpassed its predecessor. On January 20, 2025, DeepSeek-R1 debuted, excelling in mathematics, coding, and logic tasks, achieving performance levels rivaling OpenAI’s GPT models. Just a week later, the launch of the Janus Pro 7B and 1.5B models showcased their compatibility with consumer-grade hardware, thanks to their streamlined parameter sizes.
Innovative Training Techniques
What sets DeepSeek apart is its innovative training methodology. While the training costs for DeepSeek-R1 remained undisclosed, reports suggest it utilized a specialized version of Nvidia technology available exclusively in China due to U.S. export restrictions. This strategic approach significantly reduced training costs compared to other LLMs, drawing mixed reactions from industry experts, yet many commend the resourceful tactics employed by the DeepSeek team.
The Rise of DeepSeek
DeepSeek’s ascent has reverberated throughout the tech industry, prompting investors to reconsider NVIDIA’s growth strategy, resulting in a notable drop in Nvidia’s market capitalization.
The company has established a more cost-effective training paradigm, lowering barriers for AI experimentation and development. Institutions like the University of California, Berkeley, and the Hong Kong University of Science and Technology have begun small-scale experiments with DeepSeek, validating its potential and sparking interest among smaller research labs and startups eager to engage in AI development.
Prominent AI companies, including Microsoft Azure, Amazon Bedrock, Nvidia NIM, Perplexity, and Hugging Face, are now expressing support for DeepSeek-R1, accelerating the pace of innovation and real-world applications.
Despite DeepSeek’s rapid rise, insiders reveal that Liang and his team wish for the public’s attention to subside, allowing them to refocus on research and development. Established in 2023, DeepSeek has maintained a low profile, relying entirely on Liang’s self-funding without seeking external investments.
A Unique Journey from Quant Trading to Open-Source AI
Before DeepSeek, Liang’s team invested in over 10,000 GPUs for quantitative trading model training. The origins of DeepSeek are closely linked to Magic Square Quantitative, a quant fund Liang co-founded with a college friend. DeepSeek was effectively incubated within this fund, which centers its strategy around AI.
Magic Square Quantitative has been a pioneer in AI-driven quantitative trading since 2008, transitioning to a fully AI-based model by 2017. Its AI training platform, Firefly, has enhanced computational efficiency, crucial for strategy research and model testing.
Liang believes that fund managers function like programmers and servers. In a keynote titled, “The Future of Quantitative Investment in China from a Programmer’s Perspective,” he stated, “Investment decisions made by humans are often influenced by emotions, whereas those made by computer programs are grounded in scientific methodology.”
By 2019, Magic Square’s assets exceeded 10 billion RMB (1.39 billion USD), and by 2021, it became the first Chinese quantitative private equity firm to surpass 100 billion RMB (13.9 billion USD), earning recognition as one of China’s “Four Kings” of quant PE. This financial success empowered Liang to launch DeepSeek without external funding, prioritizing technological advancement.
The Philosophy Behind Open Source
Liang, who achieved his first 10 billion RMB in his 30s through technological expertise, chose an open-source model for DeepSeek. He explained, “The AI industry is still nascent, and closed-source models are challenging to commercialize quickly. An open-source ecosystem fosters developer participation, benefiting a broader audience and accelerating innovation.”
DeepSeek’s name reflects a commitment to profound exploration, with aspirations to develop artificial general intelligence (AGI) and push the boundaries of AI capabilities. The technical foundations of Magic Square and DeepSeek are intertwined, focusing on algorithm development, big data processing, and high-performance computing.
Liang’s hands-on approach is evident in his commitment to the technical aspects of DeepSeek. He is known for his unkempt appearance, yet his passion for AI development drives him to oversee even the smallest project details.
A Strong Team and Vision
DeepSeek’s team comprises top-tier talent in mathematics, physics, informatics, and AI. The recruitment strategy emphasizes collaboration across disciplines, enabling the team to tackle complex challenges in deep learning and big data modeling.
Liang asserts that a company’s true value lies in its team, especially in the face of disruptive technology. He believes that “the moat formed by closed-source models is short-lived,” emphasizing that DeepSeek’s success is rooted in its innovative culture and collaborative spirit.
In a landscape where many AI companies grapple with balancing financial pressures and technological performance, Liang’s journey stands out as a testament to the power of open collaboration in fostering a thriving AI ecosystem.
The Alchemist of AI
In China, training large language models (LLMs) is often likened to “alchemy,” a process filled with uncertainty and complexity akin to ancient legends. This reflects the journey of refining LLMs, where success is the result of extensive trials and innovation.
Liang’s story—of a hedge fund manager driven by a passion for technology and a commitment to open-source AI—highlights a higher purpose: advancing AI through collaboration and innovation.
