PeopleofSuperintelligence

Researcher

Xingwu Sun

Also known as Samm Sun

Xingwu Sun is a researcher whose work spans vision-language models, large language models, and efficient training. He coauthored LightVLP and Hunyuan-Large in 2024, and published research on optimizer scaling, Mamba-based language models, and low-precision training through 2025. In January 2026, he was a PhD candidate at the University of Macau with an oral defense scheduled for his pretrained-model thesis. His 2026 papers explored a Transformer-Mamba hybrid and mixture-of-experts routing, and described Hunyuan-A13B. As of October 2026, he is a researcher in Tencent Hunyuan's Large Language Model Department and leads pretraining for the Hunyuan model.

Now at
Tencent

Timeline

  1. May 2024

    LightVLP for vision-language pretraining

    In May 2024, he coauthors LightVLP, a vision-language pretraining method that uses two autoencoders to process unmasked tokens and reconstruct masked ones.

  2. May 2024

    The learning-rate surge

    An optimizer study he coauthors in May 2024 finds that Adam-style optimal learning rates first rise and then fall as batch size increases. Its analysis finds that the peak shifts toward larger batches as training progresses.

  3. November 2024

    The Hunyuan-Large model

    In November 2024, he coauthors the Hunyuan-Large report, which presents an open-source mixture-of-experts model with 389 billion total parameters, 52 billion activation parameters, and support for up to 256K tokens.

  4. May 2025

    Hunyuan-TurboS combines two architectures

    In May 2025, he coauthors the Hunyuan-TurboS report, which combines Mamba's long-sequence processing with Transformer contextual modeling and an adaptive long-short chain-of-thought mechanism.

  5. July 2025

    Low-precision training laws

    At ICML in July 2025, he coauthors a study of floating-point quantization for LLM training. It examines exponent and mantissa choices and estimates that the best cost-performance precision lies between 4 and 8 bits.

  6. January 2026

    A PhD defense at Macau

    At the University of Macau, he is a PhD candidate in its Department of Computer and Information Science, Faculty of Science and Technology, in January 2026. His thesis “Towards Reliable, Efficient, and Powerful Pretrained Models” is scheduled for oral defense in January 2026 under supervisor Cheng-Zhong Xu.

  7. 2026

    Union-of-Experts and TransMamba

    In March 2026, he coauthors AAAI's “TransMamba,” a sequence-level hybrid that combines Transformer and Mamba, while affiliated with Tencent's Large Language Model Department and the University of Macau. The ACL 2026 paper “Union-of-Experts” routes tokens through designated neurons inside each expert, and he publishes as Samm Sun.

  8. As of October 2026

    Hunyuan-A13B and Tencent role

    In September 2026, he coauthors the Hunyuan-A13B report, which presents an open-source mixture-of-experts model with 80 billion total parameters and 13 billion active at inference. As of October 2026, he is a researcher in Tencent Hunyuan's Large Language Model Department and leads pretraining for the Hunyuan model.

Career

  1. NowTencentResearcher and pretraining lead

Education

  1. University of MacauPhD

From the biography.

  1. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes Tencent. 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.
  2. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Tencent. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.
  3. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes Tencent. 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.
  4. The United States employs an estimated 48% of people shaping AIIncludes University of Macau. We estimate where the people shaping AI work today, how universities, companies and research institutes split across countries, and how country shares have changed since the 1960s.

Sources and credits

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

  1. LightVLP: A Lightweight Vision-Language Pre-training via Gated Interactive Masked AutoEncoders aclanthology.org. Checked 20 October 2026.
  2. Surge Phenomenon in Optimal Learning Rate and Batch Size Scaling arxiv.org. Checked 20 October 2026.
  3. Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent arxiv.org. Checked 20 October 2026.
  4. Hunyuan-TurboS: Advancing Large Language Models through Mamba-Transformer Synergy and Adaptive Chain-of-Thought arxiv.org. Checked 20 October 2026.
  5. Scaling Laws for Floating–Point Quantization Training proceedings.mlr.press. Checked 20 October 2026.
  6. PhD Oral Defense in Faculty of Science and Technology, by Mr. Xingwu SUN, on 27 January 2026 – UM E-bulletin board e-bulletin.um.edu.mo. Checked 20 October 2026.
  7. TransMamba: A Sequence-Level Hybrid Transformer-Mamba Language Model ojs.aaai.org. Checked 20 October 2026.
  8. Xingwu Sun - ACL Anthology aclanthology.org. Checked 20 October 2026.
  9. Hunyuan-A13B Technical Report arxiv.org. Checked 20 October 2026.
  10. SUN Xingwu (孙兴武) — China AI Atlas ai.techbuzzchina.com. Checked 20 October 2026.