Huawei
corporateNoah's Ark Lab is the primary AI research division of Huawei Technologies, established in 2012 as part of the Huawei 2012 Laboratories (named after the company's ethos of preparing for worst-case scenarios). Headquartered in Hong Kong with offices in Shenzhen, Beijing, Shanghai, Edmonton, London, and Paris, the lab conducts fundamental research in NLP, computer vision, recommendation systems, and AI for science.
US sanctions since 2019 have profoundly shaped Huawei's AI strategy, making domestic compute sovereignty a critical priority. Huawei's Ascend AI chip series (910, 910B, 910C) and the CANN software stack have become the backbone of China's efforts to reduce dependence on NVIDIA GPUs. Many Chinese AI labs — including Z.ai, StepFun, and state-backed institutions — now train on Ascend clusters. The MindSpore deep learning framework is Huawei's answer to TensorFlow/PyTorch.
The lab's model research centers on the PanGu series, developed in collaboration with Pengcheng Laboratory and trained on its Cloud Brain II supercomputer. PanGu-Alpha (2021, 200B) was one of China's earliest large LLMs, and PanGu-Sigma (2023, 1T sparse MoE on Ascend 910) marked a trillion-parameter milestone — but also the end of Huawei's bid to compete in the general-purpose chatbot race. After 2023, Huawei pivoted decisively to "AI for Industry" with a tiered "5+N+X" architecture: five L0 foundation models (NLP, CV, multimodal, prediction, scientific computing), industry-specific L1 models (finance, mining, meteorology), and scenario-specific L2 models. PanGu 5.5 (June 2025, 718B MoE with 256 experts) introduced deep reasoning for industrial R&D while achieving 8x better inference efficiency than PanGu-Sigma. PanGu-Weather became the first AI model to outperform traditional numerical weather prediction (published in Nature), validating Huawei's thesis that specialized models can achieve global impact without being a chatbot.
Other widely-cited research outputs include efficient architectures like GhostNet and TinyBERT, and the HEBO Bayesian optimization library which won the NeurIPS 2020 Black-Box Optimization Challenge.
People
- Tian Qi OpenReview — Former Chief Scientist (now Huawei Consumer BG)
- Zhenguo Li OpenReview — Director, AI Theory Lab