Liang Wenfeng founded DeepSeek, the AI research company associated with open model releases such as DeepSeek-V2 and R1. He appears as Wenfeng Liang on the research author lists; his published interviews explain the organization’s emphasis on original research and open release.[1][3][4][5]
Start with DeepSeek-R1 →
Background
Zhejiang University identifies Liang as a former student of its College of Information Science & Electronic Engineering and discusses his completed master’s thesis. His acknowledgement credits supervisor Xiang Zhiyu with introducing him to machine vision and guiding his research, including working through code together. This background connects his early training in visual computing with his later work on language models.[2]
Career
An official Zhejiang account identifies Liang as the founder of DeepSeek. In the published founder interviews, he explains that the team created a separate company to pursue general AI, rather than a model restricted to finance. He describes the move as a continuation of the team’s interest in difficult AI problems.[1][3]
Contributions
Making large models cheaper to run
Liang is a named co-author of DeepSeek-V2 and DeepSeek-Coder-V2. V2 combines sparse expert computation, which activates only part of a model for each token, with Multi-head Latent Attention, which compresses the stored attention information used during generation. Both changes target the resources needed to train or run a model. Coder-V2 continues training from a V2 checkpoint to strengthen coding and mathematical capabilities. These are team research results; the author lists do not assign the architecture’s invention to Liang alone.[5][6]
Reasoning and the infrastructure behind agents
The R1 paper lists Liang among its authors and studies how reinforcement learning can encourage reasoning behavior. Instead of relying entirely on human-written reasoning examples, the training process rewards performance on tasks whose answers can be checked. A later team report, DeepSeek Elastic Compute, lists him as a co-author of infrastructure that supplies isolated execution environments for agents to inspect repositories, run commands and retain state. The two projects address different layers: how a model learns and where its tool-using behavior can run.[4][7]
Ideas
In interviews originally conducted by 36Kr’s Waves team and republished by the Beijing Academy of Artificial Intelligence community, Liang argues that China should contribute original AI research. He describes open publication as a way to attract researchers and develop the team’s knowledge. These are his stated priorities at the time of the interviews, rather than a permanent guarantee about future releases. Start with the R1 paper’s abstract and training overview to connect that philosophy to a specific published experiment.[3][4]
Tools, writing & talks
Writing, research & talks
Start here
DeepSeek-R1arxiv.orgResearch on encouraging reasoning through reinforcement learning, with Liang listed as a co-author. Begin with the abstract and training overview; the experiments concern evaluated reasoning tasks rather than unrestricted real-world competence.
Connects the founder’s research philosophy to a documented team result; start with the abstract before the technical training details.
Named co-author as Wenfeng Liang[4]- DeepSeek-V2arxiv.org
A technical account of sparse expert computation and compressed attention storage. It explains why architecture affects the cost of serving language models; the detailed sections assume familiarity with neural networks.
Named co-author as Wenfeng Liang[5] - DeepSeek-Coder-V2arxiv.org
The team’s report on adapting a V2 checkpoint for programming and mathematics. It provides a concrete example of additional training for a specialized use, with benchmark claims scoped to the paper’s comparisons.
Named co-author as Wenfeng Liang[6] - DeepSeek Elastic Computearxiv.org
A 2026 report on the execution infrastructure used to train and evaluate agents at scale. It explains why tool-using models need stateful, isolated computing environments alongside model-training hardware.
Named co-author as Wenfeng Liang[7]