Wei Li

Wei Li 李韡

Not Researcher, But Engineer


韡 Biography

Since August 2025, I have been with Tencent Hunyuan, where I work on large language model (LLM) pre-training and related research engineering, contributing to the Hy3 and Hy4 model series. Previously, I was at SenseTime from August 2019 until joining Tencent, where I served as a senior algorithm researcher and led a team focused on the advancement and application of universal multimodal embedding and retrieval(UME) to real-world applications. Earlier in my career, I specialized in representation learning, implementing it in practical applications across various industries.

Before that, I earned my Master's degree in Computer Science from the Computer Games and Intelligence (CGI) Lab at National Chiao Tung University in June 2019, where my research focused on combining deep reinforcement learning with Computer Game AI (e.g., 2048 Game AI, Chinese Chess AI, AlphaGo/AlphaZero), under the supervision of Professor I-Chen Wu.

I am originally from Fujian, China, a land of rolling hills and mountains, known for its pristine natural beauty and fresh, clean air.
My name, 韡 (wěi), symbolizes brightness and beauty, reflecting my approach to both life and work. I aim to leverage technology to enhance human happiness and well-being. I'm also active on GitHub and Zhihu, feel free to connect with me!

News

2026/11 Hy4 and Hy4 flash released: 770B / 49B with vision and computer use, 224B / 12B at ~1/15 the per-task cost.

2026/07 Hy3 released and open-sourced (295B MoE / 21B active, 256K context).

2025/08 I joined Tencent Hunyuan, working on LLM pre-training and engineering.

2025/02 🎉 Our team received the SenseTime Team Award (highest group honor, with a prize of ¥500,000 CNY).

2024/05 One paper (EMAE) accepted to TPAMI'24.

2023/12 Four papers (BLV, MIAD, SpecularDefect9, SAIM) published at CVPR/ICCV/OLEN/AAAI 2023.

2022/12 Four papers (U2PL, UniVIP, Obj2Seq, FST) published at CVPR/NeurIPS 2022.

2021/09 One paper (MST) accepted to NeurIPS'21.

2020/05 😸 A Chi's (Ice Cream) broke into my life.

2019/12 👻 I received the SenseTime 2019 Outstanding Award (杰出个人奖).

2019/08 I formally joined SenseTime as a researcher.

2019/06 🌈🌈 I defended my thesis and graduated — deeply grateful to my advisor and classmates for their guidance and support, and for making these years both intellectually enriching and memorable. Our paper MPV-MCTS was accepted to IJCAI'19; it introduced a scalable AlphaGo Zero/AlphaZero (AGZ/AZ) framework and a training method that combines multiple neural networks, which helped WahahaNoGo win the TAAI 2018 and ICGA 2019 NoGo championships.

2019/03 Started an internship at SenseTime, focusing on large-scale training and face anti-spoofing.

2018/11 Worked on Project-Theta (defect inspection and automated optical inspection) and contributed to open-source projects; one paper (HTD) was accepted to WACV'19.

2017/11 👻 I received the NCTU Academic Excellence Award (書卷獎, ranked 5/267, top 1.87%).

2017/09 Served as a TA for MediaTek's Deep Learning Training Program.

2016/09 Joined the Computer Games and Intelligence (CGI) Lab at National Chiao Tung University for my master's degree, under the supervision of Professor I-Chen Wu (the inventor of Connect6).


Projects

Industrial Applications @SenseTime
  • 商汤方舟多模态新智平台 [Product]
  • 星空高铁4C智能分析系统 [News]
  • 人脸活体检测/大规模分类训练 [Product]
Efficient and Flexible Framework for AlphaZero-Like Applications @CGI Lab, NCTU
  • Developed a lightweight, scalable AGZ/AZ framework(called CFZ) on Caffe2 with C++ inference and Python training, extendable to any board game and adaptable to algorithms beyond AGZ/AZ.  CZF Guide (2MB)
  • Co-proposed a training algorithm, MPV-MCTS, based on CFZ by combining multiple NN networks.  Slides (6MB)
  • Co-developed WahahaNoGo, the successor to HaHaNoGo, with a 99% win rate against.
  • WahahaNoGo won the TAAI 2018 and ICGA 2019 NoGo Championships.
Project-Theta @NCTU, Academia Sinica
  • distribuuuu: 278
  • CIFAR-ZOO: 706
  • cifar-10-cnn: 842