In early 2025, as DeepSeek-V3 stunned the global AI community with its unprecedented training efficiency, this young AI company once again captured worldwide attention. As a subsidiary of HuanFang Technology, DeepSeek has transformed from an unknown startup to a global technology leader in just eighteen months.
HuanFang Technology: From Quantitative Trading to AI Innovation
In Hangzhou's Huijin International Building, overlooking the historic Grand Canal, a team led by 80s-born founder Liang Wenfeng is orchestrating a technological revolution. This is the birthplace of DeepSeek, emerging from quantitative trading giant HuanFang Technology.
Founded in 2016, HuanFang Technology established itself as China's leading quantitative hedge fund. Leveraging its deep expertise in AI and quantitative trading, the company began positioning itself for entry into the large language model space as early as 2022. During the severe chip shortage, HuanFang presciently secured 10,000 A100 GPUs, laying a solid foundation for DeepSeek's birth.
A Unique Technical Approach: Focus on Original Innovation
Unlike other Chinese AI companies, DeepSeek chose a distinctive development path. Founder Liang Wenfeng firmly believes that Chinese AI enterprises shouldn't remain perpetual followers but should contribute to the global wave of technological innovation.
This belief is reflected in every major company decision. Among China's seven leading AI model startups, DeepSeek stands alone in its exclusive focus on research and technology without pursuing consumer applications. It's also the only company that hasn't prioritized commercialization, instead firmly committing to open source while avoiding external funding.
Breakthrough Technical Innovations
DeepSeek's technical innovations primarily manifest in two areas: architectural innovation and training methodology breakthroughs.
Architecturally, the company's pioneering MLA (Multi-head Latent Attention) mechanism reduces memory usage to 5-13% of traditional MHA architecture. This innovation originated from a young researcher's insight and was realized through months of team effort. Additionally, the DeepSeekMoE architecture dramatically reduces computational requirements.
In training methodology, DeepSeek pioneered FP8 mixed-precision training with optimized data processing strategies. These innovations enabled DeepSeek-V3 to complete training with just 2,048 GPUs in less than two months, at a total cost of $5.5 million - approximately one-twentieth of GPT-4o's training cost.
A Young Innovation Team
DeepSeek boasts a distinctive research team. Core technical positions are primarily filled by fresh graduates and those with 1-2 years of experience, while team leaders typically have 4-6 years of experience. These young talents from top universities like Tsinghua and Peking University thrive in the company's flat management structure and innovative atmosphere.
At DeepSeek, anyone can freely access training cluster resources without approval. When a researcher has an idea, they can immediately form teams with interested colleagues. This flexible innovation mechanism has enabled numerous breakthrough ideas to materialize.
Open Source Technical Philosophy
As an innovative enterprise incubated by HuanFang Technology, DeepSeek inherited its parent company's practical innovation genes while developing its own characteristics. The company maintains an open-source approach, believing that moats formed by closed systems are temporary in the face of disruptive technology.
DeepSeek embeds value in its team, promoting the development of the entire AI ecosystem through open source. Liang Wenfeng believes that "open source is more of a cultural behavior than a commercial one. Giving is an additional honor."
Global Technical Ambitions
In an AI wave largely driven by Silicon Valley, DeepSeek provides a rare reverse case. Its MLA architecture innovation was praised by SemiAnalysis's chief analyst as "potentially the best paper this year," while former OpenAI employees lauded its "amazingly wise" training setup.
As HuanFang Technology's key AI initiative, DeepSeek is demonstrating that Chinese technology companies are fully capable of contributing at the global innovation frontier. From simple technology followers to active innovation contributors, this young company is writing a new chapter in China's AI industry.
Future Outlook
Looking ahead, DeepSeek aims to become a leading global provider of foundation models. The company hopes to form a complete industrial ecology through continuous technological innovation, enabling more enterprises to build B2B and B2C businesses on DeepSeek's foundation.
From HuanFang Technology's quantitative trading roots to its current AI frontier expansion, DeepSeek exemplifies Chinese technology companies' evolution from 1 to N, and then from 0 to 1. This story continues to unfold, and the combination of HuanFang Technology and DeepSeek will undoubtedly write a remarkable chapter in the AI era.