DeepSeek

Liang Wenfeng’s research background

Zhejiang University identifies Liang Wenfeng as DeepSeek’s founder. He completed his master’s thesis in its College of Information Science and Electronic Engineering. His supervisor, Xiang Zhiyu, introduced him to machine vision. The university describes practical research training that included close work with code.

A January 2025 notice from Hangzhou’s Information Office connects Liang’s earlier work at High-Flyer AI Fundamental Research with DeepSeek. It records his participation in Premier Li Qiang’s roundtable on 20 January. This establishes a founder’s history, not the company’s present ownership structure.

What the company’s own records establish

DeepSeek’s early Coder service agreement carries an effective date of 8 October 2023. The document names operating companies in Hangzhou and Beijing. It is evidence of activity by that date; it is not an incorporation certificate.

The English terms updated in March 2026 identify Hangzhou DeepSeek Artificial Intelligence Co., Ltd. as the owner and operator of DeepSeek products and services. The Chinese privacy policy supplies its Chinese legal name. Neither document provides a current shareholder table or an executive list.

Practical guidance. Keep the product operator, founder and investor roles separate when comparing providers. A service agreement identifies who runs the service. It does not, by itself, prove who owns every share or who holds a particular executive title.

The work extends beyond a chat model

The DeepSeek LLM paper, submitted on 5 January 2024, describes research on scaling language models and the development of 7B and 67B releases. It also describes supervised training and preference optimization used to produce chat models from the base models.

The R1 research paper followed in January 2025. It studies how reinforcement learning can develop reasoning behavior and how larger models can guide smaller ones. Liang appears in the author list as Wenfeng Liang and is listed as the paper’s submitter.

DeepSeek also publishes the Fire-Flyer File System, known as 3FS. This shared storage software supports AI training and inference. It uses SSD storage and fast networking, with documented uses in data preparation, data loading and model checkpointing. This is infrastructure software rather than another generative model.

Practical guidance. The public record shows work on both model behavior and the systems that run models. These are separate technical contributions. A software release should therefore be read as evidence of engineering scope, not as a benchmark result for a current chat API.