DeepSeek R1: An Open‑Source Large Model Matching OpenAI’s o1 at a Fraction of the Cost

On January 20, DeepSeek launched its open‑source inference model DeepSeek‑R1, which rivals OpenAI’s o1 in mathematics, programming, and reasoning tasks, yet its API calls cost 90‑95% less than competing services.

The release attracted high‑profile attention: the Chinese premier highlighted AI’s role in economic growth, Andrew Ng praised DeepSeek’s progress at the World Economic Forum, and Nvidia researcher Jim Fan called it the year’s biggest LLM underdog.

Technically, DeepSeek reduces reliance on high‑end GPUs through a Multi‑Head Latent Attention (MLA) mechanism that compresses information, and a DeepSeekMoE expert‑mix architecture that activates only a subset of experts per token, dramatically lowering memory and compute demands.

Training the model required roughly 2,048 Nvidia H800 chips—custom, lower‑spec versions of the H100—over two months for a total cost of about $5.58 million, a stark contrast to the $78 million spent on GPT‑4 and over $100 million on Llama 3.

DeepSeek adopts an open‑source model, releasing its technology freely to developers worldwide, which not only builds a strong technical reputation but also drives down overall AI model pricing across the industry.

The company originated from the Chinese quantitative fund High‑Flyer Quant, founded by Liang Wenfeng and Xu Jin. After a series of milestones—including DeepSeek‑LLM (670 B parameters), DeepSeek‑V2, and DeepSeek‑V3—the R1 model represents the latest step in their rapid development timeline.

DeepSeek’s workforce is unusually young: the team of under 140 members consists largely of recent graduates and early‑career researchers, with a hiring philosophy that favors potential over years of experience. Notable contributors include researchers such as Gao Huazuo, Zeng Wangding, Shao Zhihong, and Zhu Qihao.

Management emphasizes flat, self‑organized teams, abundant compute resources without bureaucratic approval, and no KPI pressure, fostering an environment where ideas can be quickly prototyped and scaled. This “natural division” approach is credited with accelerating innovation and keeping DeepSeek at the forefront of AI research.

Looking ahead, DeepSeek aims to shift China’s AI narrative from follower to innovator, leveraging confidence, open‑source collaboration, and cost‑effective research to compete globally.