DeepSeek CEO, Chinese open-source AI competitor
Liang Wenfeng
Profile
Liang Wenfeng is the founder and CEO of DeepSeek, the Chinese AI lab that turned the global AI industry upside down in January 2025. A quiet, unusually technical founder by Chinese tech standards, he is the rare CEO who still writes code, reads papers, and hires researchers with the same instincts as Sam Altman or Dario Amodei — except he does it from Hangzhou, on sanctioned hardware, and gives the weights away.
Born in 1985 in a small town in Guangdong, Liang studied electronic information engineering at Zhejiang University. In 2015 he co-founded High-Flyer, a quantitative hedge fund that used machine learning to trade Chinese equities. High-Flyer made him wealthy and, more importantly, gave him a reason to stockpile thousands of NVIDIA A100 GPUs before US export controls locked China out of the best silicon. In 2023 he spun that GPU hoard into DeepSeek, an AI research lab with a culture modeled openly on early OpenAI — young researchers, no KPIs, no product pressure, just training runs.
The payoff came fast. DeepSeek-V2 in mid-2024 introduced Multi-head Latent Attention and a sparse MoE architecture that slashed inference costs. DeepSeek-V3 landed in December 2024 trained for a reported $5.6M in GPU hours. Then came DeepSeek-R1 in January 2025 — an open-weight reasoning model that matched OpenAI’s o1 on math and coding benchmarks, released under an MIT license with a full technical paper. It wiped roughly $600B off NVIDIA’s market cap in a single day and forced every Western lab to explain, publicly, why their models cost so much more.
For developers learning AI, Liang’s work is the clearest demonstration that the frontier is not owned. R1 and its distillations run locally on consumer hardware, power cheap API endpoints, and underpin a growing chunk of the open-source stack. Whether you see him as a serious rival to the US labs or a strategic asset of the Chinese state — and the honest answer is probably both — DeepSeek’s weights are on Hugging Face, and they work.
Key Articles & Papers
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning DeepSeek-V3 Technical Report DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models An Interview With The Chinese CEO Behind DeepSeek DeepSeek FAQ DeepSeek-Coder: When the Large Language Model Meets ProgrammingControversies
- Training data provenance. OpenAI and Microsoft publicly suggested DeepSeek may have distilled outputs from GPT-4 to bootstrap its models. DeepSeek has not directly addressed the claim; no hard evidence has been published. Ironic, given OpenAI’s own training-data history.
- National security and CCP ties. Multiple Western governments (US, Italy, Taiwan, Australia, South Korea) have restricted or banned the DeepSeek app on government devices over data residency and censorship concerns. The hosted models refuse questions about Tiananmen, Xi Jinping, and Taiwan — the open weights, running locally, do not.
- Export-control workarounds. Reporting suggests High-Flyer’s pre-2022 A100 stockpile, plus later H800 access, is how DeepSeek trained frontier models despite sanctions. This has reopened the debate over whether US chip controls actually work.
Spotify Podcasts
![[*] Read Biography of Liang Wenfeng: The Visionary Behind DeepSeek AI and China's Tech Revolution Full](/images/people/liang-wenfeng/spotify/5n35g6iNJngWU66hBZQqe2_hu_6012275791c8bc83.webp)
