PeopleofSuperintelligence

Researcher

Hui Qu

Hui Qu is a member of the DeepSeek-AI team in Hangzhou, China, and a coauthor of DeepSeek's language-model reports from DeepSeek LLM to DeepSeek-V4.1-Flash. They coauthored DeepSeek LLM in January 2024 and DeepSeek-V2 in May 2024. In August 2024, their name appeared on the Fire-Flyer AI-HPC paper, which described a training cluster of 10,000 PCIe A100 GPUs, and in December 2024 on the DeepSeek-V3 Technical Report. DeepSeek-R1, a study of reasoning learned through pure reinforcement learning that they coauthored, reached arXiv in January 2025 and Nature in September 2025. Later reports with them among the authors include DeepSeek-V3.2 in December 2025, DeepSeek-V4 in April 2026 and DeepSeek-V4.1-Flash in September 2026.

Timeline

  1. January 2024

    The DeepSeek LLM scaling study

    In January 2024, they coauthor DeepSeek LLM, a study of scaling laws that introduces open-source models at 7B and 67B parameters. The models are pre-trained on a dataset of 2 trillion tokens, and the 67B model outperforms LLaMA-2 70B on several benchmarks, especially in code, mathematics and reasoning.

  2. May 2024

    DeepSeek-V2 and latent attention

    They are among the authors of DeepSeek-V2, released in May 2024, a Mixture-of-Experts language model with 236B total parameters and 21B activated per token. Its Multi-head Latent Attention shrinks the KV cache by 93.3%, and training costs 42.5% less than for DeepSeek 67B.

  3. August 2024

    The Fire-Flyer AI-HPC cluster

    In August 2024, they coauthor the Fire-Flyer AI-HPC paper on hardware-software co-design for deep learning. Its Fire-Flyer 2 cluster of 10,000 PCIe A100 GPUs approaches DGX-A100 performance at half the cost and with 40% less energy, using the HFReduce, HaiScale, 3FS and HAI-Platform software.

  4. December 2024

    The DeepSeek-V3 report

    December 2024 brings the DeepSeek-V3 Technical Report, which they coauthor. DeepSeek-V3 is a Mixture-of-Experts model with 671B total parameters and 37B activated per token, pre-trained on 14.8 trillion tokens and needing 2.788M H800 GPU hours for its full training.

  5. January–September 2025

    DeepSeek-R1 in Nature

    DeepSeek-R1, which they coauthor, appears on arXiv in January 2025 and shows that the reasoning abilities of language models can be incentivized through pure reinforcement learning. Nature publishes the paper in September 2025, listing Hui Qu with the DeepSeek-AI Team in Hangzhou, China.

  6. December 2025

    DeepSeek-V3.2 and sparse attention

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

  7. April 2026

    DeepSeek-V4 million-token models

    The DeepSeek-V4 preview report, which they coauthor in April 2026, presents DeepSeek-V4-Pro with 1.6T parameters and DeepSeek-V4-Flash with 284B parameters, both supporting a one-million-token context. The models combine Compressed Sparse Attention with Heavily Compressed Attention and train with the Muon optimizer on more than 32T tokens.

  8. September 2026

    The DeepSeek-V4.1-Flash model

    In September 2026, they are part of the DeepSeek-AI author team for DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model with 552B backbone parameters and contexts of up to one million tokens. Its design shrinks the global KV cache to 890 bytes per token, about a quarter of the DeepSeek-V4-Flash footprint.

  1. Google had the most people shaping AI every year from 2014 to 2025Our estimates of which organizations employed the most people shaping AI from 1961 to 2025, with a year-by-year chart.
  2. The United States employs an estimated 53% of people in our data with a known employment countryOur estimates of where AI careers are concentrated now and how countries’ shares of dated positions changed over time.
  3. 46% of living people in our data with known birth and employment countries work outside their birth countryOur estimate of cross-border careers in AI, from country of birth and the country of each current organization.
  4. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionOur estimates of where the people shaping AI studied and where their careers led, from education records in cited biographies.

Sources and credits

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

  1. [2401.02954] DeepSeek LLM: Scaling Open-Source Language Models with Longtermism arxiv.org. Checked 19 October 2026.
  2. [2405.04434] DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model arxiv.org. Checked 19 October 2026.
  3. [2408.14158] Fire-Flyer AI-HPC: A Cost-Effective Software-Hardware Co-Design for Deep Learning arxiv.org. Checked 19 October 2026.
  4. [2412.19437] DeepSeek-V3 Technical Report arxiv.org. Checked 19 October 2026.
  5. [2501.12948] DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning arxiv.org. Checked 19 October 2026.
  6. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning | Nature nature.com. Checked 19 October 2026.
  7. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning - PMC pmc.ncbi.nlm.nih.gov. Checked 19 October 2026.
  8. [2512.02556] DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models arxiv.org. Checked 19 October 2026.
  9. [2606.19348] DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence arxiv.org. Checked 19 October 2026.
  10. [2609.19969] DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression arxiv.org. Checked 19 October 2026.