PeopleofSuperintelligence

Researcher

Zhengyan Zhang

Zhengyan Zhang is a language-model researcher whose work runs from knowledge-enhanced pre-training at Tsinghua University to large-scale model building at DeepSeek. Supervised by Zhiyuan Liu in Tsinghua's Department of Computer Science and Technology, they worked as a PhD student on pre-trained models for natural language processing. ERNIE appeared in 2019 and CPM in 2020 with them as first-listed author, and MoEfication, which turns Transformer feed-forward layers into mixtures of experts, followed in 2022. In 2024 they were named among the Research & Engineering contributors to DeepSeek-V3, and Native Sparse Attention, coauthored with Wenfeng Liang and others, won an ACL 2025 Best Paper Award. As of October 2026, they are an author of DeepSeek's V4 and V4.1-Flash reports; Nature's 2025 DeepSeek-R1 paper listed them with the DeepSeek-AI Team in Hangzhou.

Now at
DeepSeek

Timeline

  1. 2018–2021

    A survey of graph neural networks

    In December 2018, they coauthor a graph neural network survey with Jie Zhou, Ganqu Cui, Shengding Hu, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li and Maosong Sun. The review proposes a general design pipeline for GNN models and appears in AI Open in 2021.

  2. 2019

    ERNIE and knowledge graphs

    In 2019, they are first-listed author of ERNIE, a language representation model trained on large text corpora and knowledge graphs, with Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun and Qun Liu. The paper appears at the 57th Annual Meeting of the Association for Computational Linguistics in Florence.

  3. 2020

    CPM, a Chinese language model

    In December 2020, they head the author list of CPM, a generative Chinese pre-trained language model with 2.6 billion parameters trained on 100GB of Chinese data. Zhiyuan Liu, Minlie Huang and Maosong Sun are among the other authors.

  4. 2022

    MoEfication of Transformers

    In 2022, they publish MoEfication in Findings of ACL 2022 with Yankai Lin, Zhiyuan Liu, Peng Li, Maosong Sun and Jie Zhou. The method converts a Transformer into a Mixture-of-Experts version with the same parameters and uses 10% to 30% of feed-forward parameters while keeping over 95% of original performance.

  5. By 2024

    PhD at Tsinghua University

    By 2024, they are a PhD student in the Department of Computer Science and Technology at Tsinghua University, supervised by Zhiyuan Liu. Their research interests lie in pre-trained models for natural language processing.

  6. December 2024

    Contributing to DeepSeek-V3

    In December 2024, the DeepSeek-V3 Technical Report appears, describing a Mixture-of-Experts language model with 671B total parameters and 37B activated for each token. They are named among its Research & Engineering contributors.

  7. 2025

    Best paper for Native Sparse Attention

    In February 2025, they coauthor Native Sparse Attention with Jingyang Yuan, Huazuo Gao, Damai Dai, Wenfeng Liang and others. The method combines coarse-grained token compression with fine-grained token selection for long-context modeling, and it receives a Best Paper Award at ACL 2025 in Vienna.

  8. As of October 2026

    Part of the DeepSeek-AI team

    In September 2025, Nature publishes the DeepSeek-R1 paper, which lists them with the DeepSeek-AI Team in Hangzhou. They are also an author of the DeepSeek-V4 report of April 2026, which presents two Mixture-of-Experts models with a context length of one million tokens, and of the DeepSeek-V4.1-Flash report of September 2026.

Career

  1. Since 2025DeepSeek

Education

  1. Tsinghua UniversityPhD

From the biography.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes DeepSeek. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  2. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes DeepSeek. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.
  3. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staffIncludes DeepSeek. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staff. We test the estimate against 3,913 named model contributors and dated all-staff records across Google DeepMind, OpenAI and Anthropic, plus six additional frontier-grade labs.
  4. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes DeepSeek. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Zhengyan Zhang yet. Send us a free-licensed photo.

  1. Graph Neural Networks: A Review of Methods and Applications arxiv.org. Checked 21 October 2026.
  2. ERNIE: Enhanced Language Representation with Informative Entities arxiv.org. Checked 21 October 2026.
  3. ERNIE: Enhanced Language Representation with Informative Entities - ACL Anthology aclanthology.org. Checked 21 October 2026.
  4. CPM: A Large-scale Generative Chinese Pre-trained Language Model arxiv.org. Checked 21 October 2026.
  5. MoEfication: Transformer Feed-forward Layers are Mixtures of Experts - ACL Anthology aclanthology.org. Checked 21 October 2026.
  6. About me - Zhengyan Zhang zzy14.github.io. Checked 21 October 2026.
  7. DeepSeek-V3 Technical Report arxiv.org. Checked 21 October 2026.
  8. DeepSeek-V3 Technical Report (HTML version) arxiv.org. Checked 21 October 2026.
  9. Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention arxiv.org. Checked 21 October 2026.
  10. Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention - ACL Anthology aclanthology.org. Checked 21 October 2026.
  11. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning | Nature nature.com. Checked 21 October 2026.
  12. DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence arxiv.org. Checked 21 October 2026.
  13. DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression arxiv.org. Checked 21 October 2026.