Wang Yunhe

wiki
, ,
1

Former Director of Huawei Noah’s Ark Lab, core lead of the Pangu Foundation Model, deep learning researcher excelling in both model compression theory and large-scale industrial deployment, now Co-founder and CEO of AI Agent startup Jiyuan Lvdong.
Born 1991, China
Nationality Chinese
Affiliations Jiyuan Lvdong (Co-founder & CEO, Apr 2026–); formerly Huawei Noah’s Ark Lab (Director, Mar 2025–Mar 2026)
Research Deep learning, model compression and quantization, neural architecture search (NAS), computer vision, large language models, AI Agents
PhD Advisors Chao Xu (Peking University), Dacheng Tao (Fellow of the Australian Academy of Science)
Website wangyunhe.site
Zhihu YunheWang
GitHub YunheWang
Google Scholar Wang Yunhe — 33k+ citations, h-index 68

Biography

Wang Yunhe is a representative Chinese deep learning researcher known for balancing academic output with industrial deployment, specializing in efficient neural network architectures and model compression. During nearly nine years at Huawei’s Noah’s Ark Lab, he rose from intern to lab director, leading the development of widely adopted lightweight architectures like GhostNet and AdderNet, and serving as the core lead for Huawei’s Pangu Foundation Model, overseeing the Pangu 5.5 series. After leaving Huawei in March 2026, he founded Jiyuan Lvdong, a company focused on enterprise-level AI agent deployment, achieving a $100 million valuation after its initial funding round. With over 33,000 Google Scholar citations and an h-index of 68, he belongs to the top tier of industrial researchers in China.

Education and Early Life

Wang Yunhe earned his bachelor’s degree in Mathematics and Applied Mathematics from Xidian University, building a solid mathematical foundation. He then pursued a PhD at the Department of Intelligence Science, Peking University, under the supervision of Professor Chao Xu and Professor Dacheng Tao (Fellow of the Australian Academy of Sciences), focusing on deep learning and model compression.

During his PhD, he began publishing at top-tier academic conferences, including NeurIPS 2016 with CNNpack (frequency-domain neural network compression) and ICML 2017 with Beyond Filters (compact feature maps), demonstrating a research style that highly integrates frequency-domain theory with deep learning engineering. In 2017, before graduating, he joined Huawei as the first intern at the Beijing branch of the Noah’s Ark Lab, starting a dual-track career in academia and industry. He completed his PhD in 2018 and officially joined Huawei.

During his studies, he received the Google PhD Fellowship (2017), the Baidu Scholarship (2017), the Peking University President’s Scholarship (2017), and two National Postgraduate Scholarships (2016, 2017), making him one of the most highly awarded PhD students of his cohort.

Career

Huawei Noah’s Ark Lab (2017–2026)

Wang Yunhe joined Huawei’s Noah’s Ark Lab as an intern in 2017, becoming the first intern researcher recruited during the early establishment of its Beijing branch. After earning his PhD in 2018, he officially joined full-time, starting as a senior engineer and steadily rising through the ranks as principal engineer and technical expert, with a promotion pace often described as “one promotion per year” within the lab.

Efficient Deep Learning Architecture Research (2018–2021)

During this phase, Wang co-led several landmark foundational architecture studies at Huawei’s Noah’s Ark Lab. GhostNet (CVPR 2020) proposed generating “ghost feature maps” using cheap linear operations, significantly reducing parameters and computation for on-device models while maintaining accuracy. It quickly became one of the industry’s most cited lightweight architectures, with over 6,000 citations on his Google Scholar. In the same year, AdderNet (CVPR 2020 Oral) replaced multiplication with addition operations in deep networks, redefining efficient AI from a hardware power consumption perspective. These two achievements jointly won Huawei’s Fourth Top Ten Invention Award and were algorithmically optimized for deployment on China’s Sky Eye FAST telescope, assisting researchers in efficiently identifying hundreds of new fast radio burst (FRB) samples.

Additionally, the industrial application of Vision Transformers was a key focus during this period. Wang and his team published works like Transformer-in-Transformer (TNT, NeurIPS 2021), Pre-trained Image Processing Transformer (IPT, CVPR 2021), and authored the highly cited survey “A Survey on Visual Transformer” (IEEE TPAMI 2022), systematically outlining the development of ViT research.

Head of Algorithm Application Department (Late 2021–Early 2025)

At the end of 2021, Wang was promoted to Head of the Algorithm Application Department (later renamed the Small Model Lab), taking full responsibility for the R&D and deployment of efficient AI algorithms. In this role, he continued to advance the transfer of model lightweighting technology to end products, with related outcomes applied in Huawei’s flagship phones like the Mate 30 and various industry solutions.

Director of Huawei Noah’s Ark Lab & Head of Pangu Foundation Model (Mar 2025–Mar 2026)

In March 2025, following an internal transfer of the previous Noah’s Ark Lab Director, Yao Jun, Wang took on the role of lab director and concurrently became the core lead for the Pangu Foundation Model, making him the top technical leader of Huawei’s AI R&D system, known in the industry as the “Pangu Young Marshal.”

Within about three months of taking over, he rapidly drove the release of the Pangu 5.5 series. Pangu 5.5 adopts a Mixture of Experts (MoE) architecture; its flagship version, “Pangu Ultra,” has 718B parameters with about 39B activated parameters. It introduces an adaptive fast-slow thinking mechanism, achieving an 8x improvement in overall inference efficiency over its predecessor. It also features a deep retrieval Agent capability module called DeepDiver, capable of completing complex question answering involving over ten hops within 5 minutes and generating professional reports of tens of thousands of words. During this period, the Pangu Foundation Model was deeply implemented in major automotive companies including FAW, GAC, Changan, SERES, and BYD, covering the entire chain of vehicle R&D, production, and after-sales.

On March 28, 2026, Wang confirmed his departure from Huawei in a social media post, bidding farewell: „Eight years, or more precisely nine years (the first intern in Beijing in 2017). It is with reluctance that I leave this place where I once worked hard.“

Jiyuan Lvdong (Apr 2026–)

On April 8, 2026, Shanghai Jiyuan Lvdong Technology Co., Ltd. was officially registered, followed by Beijing Jiyuan Lvdong Technology Co., Ltd. on April 17, with Wang Yunhe as the legal representative of both. Former Huawei Noah’s Ark Lab Chief Researcher Kai Han joined as CTO; having long collaborated with Wang on algorithms like GhostNet, the two share a high degree of synergy in technical direction and engineering implementation.

Jiyuan Lvdong’s core track is industry-deployed enterprise-level AI Agents. Leveraging Wang’s expertise in lightweight large models and low-power algorithms accumulated at Huawei, the company focuses on helping traditional industries like manufacturing, finance, and automotive build proprietary AI assistants. It aims to address pain points such as high costs, difficult deployment, and significant computational power consumption in enterprise AI implementation, deliberately avoiding the red ocean of competition for general-purpose large models.

On June 2, 2026, Sina Tech exclusively reported that Jiyuan Lvdong had completed a funding round with a post-investment valuation of $100 million. The investors consist of first-tier domestic venture capital institutions and several leading internet companies. The company has also secured stable cooperation orders from several state-owned enterprise giants, with its first commercial Agent product planned for release within a few months.

Major Contributions

  • GhostNet (CVPR 2020) — Proposed a lightweight network architecture using cheap linear operations to generate „ghost feature maps.“ It significantly reduces parameters and FLOPs compared to previous best solutions while maintaining accuracy, becoming one of the mainstream techniques for on-device AI. With over 6,000 citations, it is Wang’s most cited paper.
  • AdderNet (CVPR 2020 Oral) — Replaces multiplication convolutions with pure addition operations in deep networks, drastically reducing computational power consumption and circuit area at the operator level. It won Huawei’s Top Ten Invention Award and was deployed on China’s Sky Eye FAST, aiding in discovering hundreds of new fast radio bursts.
  • VanillaNet (2023) — Embodies a „less is more“ minimalist design philosophy, building efficient backbone networks by reducing depth and complex operations (e.g., self-attention). A 6-layer VanillaNet surpasses ResNet-34, and a 13-layer version achieves approximately 83% Top-1 accuracy, offering significant hardware efficiency advantages.
  • PanGu-π / Pangu 5.5 (2023–2025) — Led architectural innovations for Huawei’s Pangu foundation model. PanGu-π significantly enhanced the expressiveness of 7B/1B scale LLMs through nonlinear compensation modules. Pangu 5.5 adopted an MoE architecture to achieve an 8x improvement in inference efficiency, driving the commercial deployment of large models in automotive, manufacturing, and finance industries.
  • Vision Transformer System Research — Co-authored works including TNT (NeurIPS 2021), IPT (CVPR 2021), CMT (CVPR 2022), and led the writing of the ViT survey (IEEE TPAMI 2022), one of the earliest systematic efforts to map the vision Transformer research landscape in China, with sustained high citation counts.
  • Systematic Model Quantization and Pruning Methods — Published multiple works at top conferences like NeurIPS, ICML, CVPR, ICCV on quantization (AdderQuant, low-bit search), pruning (SCOP, ManiDP), and knowledge distillation (data-free distillation DAFL, decoupled feature distillation), systematically advancing the foundational methodology of efficient AI.

Social Impact and Honors

  • Google PhD Fellowship (2017) — Very few Chinese graduate students received this fellowship that year.
  • Baidu Scholarship (2017) — Awarded annually to only about 10 top PhD students nationwide in China.
  • Peking University President’s Scholarship (2017) — One of the highest honors for graduate students at Peking University.
  • National Postgraduate Scholarship (2016, 2017) — Won for two consecutive years.
  • Huawei Fourth Top Ten Invention Award — Won for „High-Performance Multiplier and Adder Neural Network,“ Huawei’s highest-level innovation and invention award, recognizing achievements with both high academic impact and commercial value within the company.
  • WAIC Outstanding Youth Paper Award Nomination (2020) — Nominated for the best youth paper at the World Artificial Intelligence Conference.
  • CVPR 2024 Best Student Paper Nomination — His team’s paper became a candidate for the highest honor at this top international computer vision conference.

Academic and Professional Network

  • Chao Xu — PhD advisor at Peking University; Professor at the Department of Intelligent Science, a seasoned scholar in model compression and efficient deep learning, who laid the academic foundation for Wang’s research direction.
  • Dacheng Tao — Co-advisor during PhD at Peking University; Fellow of the Australian Academy of Science, an international authority in statistical machine learning and computer vision, with ongoing collaborations in later Huawei papers.
  • Kai Han — Former Chief Researcher at Huawei Noah’s Ark Lab, co-author of core papers like GhostNet and TNT; currently CTO of Jiyuan Lvdong, Wang’s long-term most important collaborator and entrepreneurial partner.
  • Qi Tian — Collaborator on works including GhostNet and AdderNet, senior researcher at Huawei Noah’s Ark Lab.
  • Hanting Chen — Core co-author of multiple top conference papers like AdderNet and IPT, researcher at Huawei Noah’s Ark Lab.
  • Chang Xu — Professor at Nanyang Technological University, academic collaborator during Wang’s PhD, with whom he co-authored several high-impact papers on model compression.

Personal Style

Wang Yunhe’s research approach exhibits a distinct style of “mathematical intuition driving engineering deployment.” The inspiration for AdderNet stemmed from a reflection on the nature of multiply-add operations during a mountain hike, while GhostNet was derived from the statistical observation of “redundant feature maps,” directly yielding a quantifiable engineering solution. Both demonstrate an ability to leap from simple mathematical insights to scalable deployment schemes. On platforms like Zhihu, he maintains open communication with academic peers and engineers, with direct technical expression and minimal embellishment. His journey from Huawei’s first Beijing intern to lab director in nine years, and then to founding a startup valued at $100 million, reflects a typical trajectory for a new generation of AI leaders in China, making independent choices between deep technical accumulation within a large corporation and the opportunity of entrepreneurship.

References