Liang Wenfeng is a Chinese entrepreneur and founder of DeepSeek, an AI research lab that drew international attention in early 2025 after releasing a reasoning model that matched leading Western systems at a small fraction of the reported training cost.
Liang trained in electronic and information engineering and information and communication engineering in China before co-founding High-Flyer, a quantitative hedge fund that used machine learning techniques for automated trading. High-Flyer's early investment in GPU infrastructure for its trading models gave Liang's team unusually strong in-house computing resources and engineering talent, which he redirected toward general AI research when he founded DeepSeek in 2023 as a spin-off effort funded by the fund.
DeepSeek-R1 and the market shock
DeepSeek released a series of open-weight large language models through 2023 and 2024, but the company's profile changed dramatically in January 2025 with the release of DeepSeek-R1, a reasoning model trained with large-scale reinforcement learning that performed competitively with OpenAI's o1 on math and coding benchmarks. DeepSeek published a technical report describing a training cost far lower than industry estimates for comparable frontier models, and released the model's weights openly, allowing developers worldwide to download and run it. The announcement triggered the DeepSeek shock, a sharp sell-off in AI-related stocks, including NVIDIA, as investors questioned assumptions about how much computing infrastructure spending frontier AI progress actually required.
Technical approach
DeepSeek's models under Liang made extensive use of mixture-of-experts architectures and efficiency-focused training techniques to reduce compute costs, and the company published unusually detailed technical papers relative to Western labs, which increasingly disclosed less about their training methods as competition intensified. Liang has given few public interviews, but in the ones he has given he described DeepSeek's mission as pursuing general AI capability research for its own sake rather than short-term commercial products.
Reception
Liang's low public profile, in contrast to the prominence of Western AI executives, drew significant media curiosity following DeepSeek-R1's release, and Chinese state media and officials highlighted the company as evidence of China's AI capability despite U.S. export controls on advanced chips. Western commentators debated whether DeepSeek's reported training costs were fully representative of total investment, while broadly agreeing the release had accelerated interest in efficient training and inference techniques industry-wide.