PeopleofSuperintelligence

Researcher

Kuai Yu

Kuai Yu is a DeepSeek-AI researcher whose work links large language models with the systems used to train and serve them. In December 2024, they were credited in DeepSeek-V3's Research & Engineering group. Their later work spans reinforcement learning, sparse attention, manifold-constrained residual connections, and communication libraries for expert-parallel and Ascend systems. As of October 2026, they work in DeepSeek-AI's Research & Engineering group and contribute to DeepSeek-V4.1-Flash, a multimodal model that supports one-million-token contexts.

Now at
DeepSeek
Works
1
Citations
0
h-index
0

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

  1. December 2024

    DeepSeek-V3 technical report

    In December 2024, the report credits Kuai Yu in its Research & Engineering group. It describes a Mixture-of-Experts language model with 671B total parameters and 37B activated for each token.

  2. January 2025–September 2025

    DeepSeek-R1 and reinforcement learning

    In January 2025, they coauthor DeepSeek-R1, whose results show that pure reinforcement learning can incentivize reasoning in large language models without human-labeled reasoning trajectories. Nature publishes the paper in September 2025 with the DeepSeek-AI Team in Hangzhou, China.

  3. 2025

    DeepEP communication library

    In 2025, they coauthor DeepEP, an expert-parallel communication library.

  4. December 2025

    DeepSeek-V3.2 and sparse attention

    In December 2025, they coauthor DeepSeek-V3.2, which introduces DeepSeek Sparse Attention to reduce computational complexity while preserving performance in long-context scenarios. Its high-compute variant, DeepSeek-V3.2-Speciale, achieves gold-medal performance at the 2025 International Mathematical Olympiad and the International Olympiad in Informatics.

  5. December 2025

    Manifold-constrained hyper-connections

    In December 2025, they join a 20-author mHC paper led by Zhenda Xie and including Wenfeng Liang. The framework projects the residual connection space of hyper-connections onto a manifold to restore identity mapping, and experiments show it is effective for training at scale.

  6. April 2026

    DeepSeek-V4 model preview

    In April 2026, they coauthor the DeepSeek-V4 preview: V4-Pro has 1.6T parameters and V4-Flash has 284B, with both supporting one-million-token contexts. The series uses Manifold-Constrained Hyper-Connections to strengthen conventional residual connections.

  7. 2026

    DeepEP-Ascend for Huawei NPUs

    In 2026, they coauthor DeepEP-Ascend, a communication library for machine-learning training and inference on Huawei Ascend NPUs.

  8. As of October 2026

    Research & Engineering at DeepSeek

    In September 2026, they coauthor DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model with 552B backbone parameters and contexts of up to one million tokens. As of October 2026, they work in DeepSeek-AI's Research & Engineering group.

Most-cited works

  1. mHC: Manifold-Constrained Hyper-Connections2025 · 0 citationsWith Chengqi Deng, Damai Dai, Huazuo Gao, Shangyan Zhou and Zhenda Xie

From OpenAlex, as of 21 October 2026.

  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 Kuai Yu yet. Send us a free-licensed photo.

  1. [2412.19437] DeepSeek-V3 Technical Report arxiv.org. Checked 21 October 2026.
  2. DeepSeek-V3 Technical Report arxiv.org. Checked 21 October 2026.
  3. [2501.12948] DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning arxiv.org. Checked 21 October 2026.
  4. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning nature.com. Checked 21 October 2026.
  5. DeepEP: an efficient expert-parallel communication library github.com. Checked 21 October 2026.
  6. DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models arxiv.org. Checked 21 October 2026.
  7. [2512.24880] mHC: Manifold-Constrained Hyper-Connections arxiv.org. Checked 21 October 2026.
  8. DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence arxiv.org. Checked 21 October 2026.
  9. A high-performance communication library for machine learning training and inference on Huawei Ascend NPUs github.com. Checked 21 October 2026.
  10. [2609.19969] DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression arxiv.org. Checked 21 October 2026.
  11. DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression arxiv.org. Checked 21 October 2026.