DeepSeek has gained widespread attention.
Recently, DeepSeek, an AI lab based in China, has made a remarkable mark in the global technology landscape as its chatbot application soared to the top of the Apple App Store charts. The AI models crafted by DeepSeek, through efficient computational strategies, have captured the interest of analysts on Wall Street and tech experts, leading to discussions regarding the United States’ capability to retain its position in the AI sector and whether the demand for AI chip technology will persist.
So, what is the background of DeepSeek, and how did it rise to global prominence so rapidly?
The Origins of DeepSeek
DeepSeek is financially supported by High-Flyer Capital Management, a quantitative hedge fund in China that employs AI to inform its trading strategies.
Losing enthusiasts Liang Wenfeng co-founded High-Flyer in 2015. Wenfeng began his trading journey while studying at Zhejiang University and later established High-Flyer Capital Management as a hedge fund in 2019, concentrating on the creation and use of AI algorithms.
In 2023, High-Flyer launched DeepSeek as a dedicated laboratory for AI research, separate from its financial operations. With High-Flyer as an investor, the lab transitioned into an independent entity, also named DeepSeek.
DeepSeek’s technical workforce is relatively youthful. Reports indicate that the company actively recruits PhD-level AI researchers from elite Chinese universities. Additionally, DeepSeek also employs individuals from various academic backgrounds to enhance its technological understanding across a broad spectrum of fields, as reported by The New York Times.
DeepSeek’s Powerful Models
In November 2023, DeepSeek introduced its initial range of models – including DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat. However, the AI community began to take serious note during the launch of its next-generation models, the DeepSeek-V2 family, in the spring of the following year.
DeepSeek-V2 is an all-purpose system adept at analysing both text and images, delivering impressive results across various AI benchmark tests and operating at a significantly lower cost than its contemporaries at the time. This competitive edge compelled domestic rivals, such as ByteDance and Alibaba, to lower their pricing on certain models and offer others at no cost.
DeepSeek-V3, which debuted in December 2024, further enhanced DeepSeek’s reputation.
According to DeepSeek’s internal performance evaluations, DeepSeek V3 surpasses both openly available models like Meta’s Llama and “closed” models that require API access, such as OpenAI’s GPT-4o.
DeepSeek’s R1 “reasoning” model, launched in January, boasts performance on par with OpenAI’s o1 model on key tests.
The R1 model can effectively verify its own facts, making it less prone to typical errors commonly encountered by other models. While reasoning models may take longer—often several seconds to minutes longer—to produce solutions compared to standard models, their increased reliability in fields such as physics, science, and mathematics is a significant advantage.
However, there is a drawback to DeepSeek V3, R1, and the other models. Developed in China, they must undergo scrutiny by China’s internet regulatory body to ensure that their responses align with socialist principles. For instance, in DeepSeek’s chatbot, the R1 model avoids responding to inquiries about Tiananmen Square or Taiwan’s independence.
A Disruptive Strategy
It remains ambiguous if DeepSeek follows a clear business model. The company sets its pricing significantly lower than the market rate and also offers numerous services at no charge.
According to DeepSeek, significant advancements in efficiency have allowed it to maintain remarkable cost competitiveness. Some specialists contest the validity of the financial data provided by the company.
Regardless of the situation, developers have shown enthusiasm for DeepSeek’s models, which, while not open source in the traditional sense, are available under flexible licenses permitting commercial applications. Clem Delangue, CEO of Hugging Face—one of the hosts of DeepSeek’s models—indicated that developers on Hugging Face have created more than 500 “derivative” models of R1, amounting to a total of 2.5 million downloads collectively.
DeepSeek’s achievement in overcoming larger, established competitors has been described as “transforming AI” and “beginning a new phase of AI competition.” The company’s rise has significantly contributed to an 18% drop in Nvidia’s stock price and prompted a public reaction from OpenAI’s CEO, Sam Altman.
The future of DeepSeek remains uncertain. Enhanced models are anticipated, but there appears to be a growing concern from the U.S. government regarding what it regards as potentially adverse foreign influence.
