PeopleofSuperintelligence

Researcher

Zhangli Sha

Zhangli Sha is an author on the technical papers behind DeepSeek's open language models, from DeepSeek LLM and DeepSeek-V3 to the reasoning model DeepSeek-R1. They coauthored DeepSeek LLM in January 2024, the Mixture-of-Experts model DeepSeek-V2 in May 2024, and the Fire-Flyer AI-HPC paper on training with 10,000 PCIe A100 GPUs in August 2024. The DeepSeek-V3 Technical Report followed in December 2024, and in September 2025 Nature published the DeepSeek-R1 paper, which listed them under the DeepSeek-AI Team in Hangzhou, China. DeepSeek-V3.2 in December 2025 and the million-token DeepSeek-V4 in April 2026 came next. They are an author of the September 2026 reports on DeepSeek-V4.1-Flash and the DSec sandbox platform for agentic training.

Timeline

  1. January 2024

    The DeepSeek LLM paper

    Zhangli Sha is an author of DeepSeek LLM: Scaling Open-Source Language Models with Longtermism, submitted in January 2024. The paper studies scaling laws and presents 7B and 67B models, and DeepSeek LLM 67B surpasses LLaMA-2 70B on benchmarks in code, mathematics and reasoning.

  2. May 2024

    DeepSeek-V2 language model

    They coauthor DeepSeek-V2 in May 2024, a Mixture-of-Experts language model with 236B total parameters and 21B activated for each token. Compared with DeepSeek 67B, the model saves 42.5% of training costs.

  3. August 2024

    Fire-Flyer AI-HPC architecture

    In August 2024, the Fire-Flyer AI-HPC paper on software-hardware co-design for deep learning appears with them as a coauthor. Its Fire-Flyer 2 deployment of 10,000 PCIe A100 GPUs approaches DGX-A100 performance at half the cost and with 40% less energy.

  4. December 2024

    The DeepSeek-V3 report

    They are among the authors of the DeepSeek-V3 Technical Report, submitted in December 2024. The model has 671B total parameters, 37B of them activated for each token, and needs 2.788M H800 GPU hours for full training.

  5. September 2025

    DeepSeek-R1 in Nature

    Nature publishes the DeepSeek-R1 paper in September 2025, with them listed under the DeepSeek-AI Team in Hangzhou, China. The paper shows that the reasoning abilities of large language models can be incentivized through pure reinforcement learning, removing the need for human-labelled reasoning trajectories.

  6. December 2025

    DeepSeek-V3.2 and sparse attention

    DeepSeek-V3.2, submitted in December 2025, lists them among its authors and introduces DeepSeek Sparse Attention for long-context efficiency. Its high-compute variant, DeepSeek-V3.2-Speciale, achieves gold-medal performance in both the 2025 International Mathematical Olympiad and the International Olympiad in Informatics.

  7. April 2026

    DeepSeek-V4 million-token models

    The DeepSeek-V4 preview, submitted in April 2026, carries their name among its authors. The series includes DeepSeek-V4-Pro with 1.6T parameters and DeepSeek-V4-Flash with 284B parameters, and both support a context length of one million tokens.

  8. September 2026

    DeepSeek-V4.1-Flash and DSec

    Two reports submitted in September 2026 carry their name: DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model with 552B backbone parameters and support for contexts of up to one million tokens, and DeepSeek Elastic Compute (DSec), a sandbox platform for agentic training that supports over 380,000 concurrent sandboxes in production.

  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.

Sources and credits

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

  1. [2401.02954] DeepSeek LLM: Scaling Open-Source Language Models with Longtermism arxiv.org. Checked 21 October 2026.
  2. [2405.04434] DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model arxiv.org. Checked 21 October 2026.
  3. [2408.14158] Fire-Flyer AI-HPC: A Cost-Effective Software-Hardware Co-Design for Deep Learning arxiv.org. Checked 21 October 2026.
  4. [2412.19437] DeepSeek-V3 Technical Report arxiv.org. Checked 21 October 2026.
  5. DeepSeek-V3 Technical Report | European Laboratory for Learning and Intelligent Systems ellis.eu. Checked 21 October 2026.
  6. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning | Nature nature.com. Checked 21 October 2026.
  7. [2512.02556] DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models arxiv.org. Checked 21 October 2026.
  8. [2606.19348] DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence arxiv.org. Checked 21 October 2026.
  9. [2609.19969] DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression arxiv.org. Checked 21 October 2026.
  10. [2609.22978] DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale arxiv.org. Checked 21 October 2026.