Huawei AI is the artificial intelligence division of Huawei Technologies Co., Ltd., a Chinese multinational technology conglomerate headquartered in Longgang, Shenzhen, Guangdong. It focuses on developing AI hardware and software, including the Ascend series of AI chips and the Pangu family of large language models, to serve enterprise, cloud, and consumer applications. The division operates within Huawei's broader strategy of building a full-stack AI ecosystem, from silicon to frameworks to model deployment, particularly in response to global supply chain restrictions and the push for technological self-reliance in China.
The company's AI efforts are part of its expansion beyond telecommunications equipment, which remains its largest business area. Founded in 1987 by Ren Zhengfei, a former deputy director of the People's Liberation Army engineering corps, Huawei has grown into a global player with operations in over 170 countries. As of 2025, it is the largest smartphone vendor in China with an 18.1% market share, and its AI initiatives are closely tied to its consumer devices, cloud services, and enterprise solutions.
Ascend AI Chips
The Ascend series is Huawei's line of AI accelerators designed for training and inference of neural networks. The first generation, Ascend 310 and Ascend 910, were announced in 2018 and targeted edge and data center applications, respectively. The Ascend 910, launched in 2019, was positioned as a high-performance training chip, with Huawei claiming it delivered up to 256 teraflops of half-precision performance. The Ascend 310, for inference, offered 8 teraflops and was aimed at low-power edge devices.
In 2023, Huawei introduced the Ascend 910B, an upgraded version with improved memory bandwidth and interconnect, designed to compete with NVIDIA's data center GPUs. The Ascend 910C followed in 2024, featuring enhanced performance and support for larger models. These chips are built on TSMC's 7-nanometer process, though US sanctions have limited access to advanced fabrication, prompting Huawei to rely on domestic supply chains. The Ascend series is integrated into Huawei's Atlas computing platforms, which are used in cloud data centers and enterprise AI clusters.
The Ascend architecture includes a dedicated AI core with a matrix multiplication unit, vector unit, and scalar unit, optimized for Deep learning workloads. It supports common frameworks like TensorFlow and PyTorch through Huawei's mindspore framework and the CANN (Compute Architecture for Neural Networks) toolkit. As of 2025, Huawei claims Ascend has been deployed in over 50 industries, including finance, healthcare, and autonomous driving.
Pangu Foundation Models
Pangu is Huawei's family of large language models and foundation models, first unveiled in 2021. The initial Pangu-α model, with 200 billion parameters, was trained on Chinese and English text and demonstrated strong performance on natural language processing tasks. Subsequent versions expanded to multimodal capabilities, including vision and speech, and specialized models for industries like medicine, mining, and weather forecasting.
In 2023, Huawei released Pangu-3.0, which introduced a modular architecture with a base model and task-specific adapters. This version emphasized efficiency, allowing fine-tuning with fewer computational resources. The Pangu models are trained on Huawei's ascend clusters and are available through Huawei Cloud, offering APIs for Generative AI applications such as text generation, code completion, and question answering.
Pangu has been integrated into Huawei's consumer products, including smartphones and smart assistants, and into enterprise solutions for Artificial intelligence-driven decision making. In 2024, Huawei announced Pangu-4.0, which improved reasoning and multilingual capabilities, and launched a lightweight version for edge devices. The models are designed to comply with Chinese regulations on AI, including content safety and data privacy.
AI Software Stack
Huawei AI is built on a software stack that includes the MindSpore framework, which is an open-source Machine learning framework similar to TensorFlow and PyTorch. MindSpore supports automatic differentiation, parallel training, and deployment across Ascend, GPU, and CPU environments. It was released in 2020 and has gained adoption in Chinese universities and enterprises.
The CANN toolkit provides low-level libraries for Ascend hardware, enabling optimized execution of neural networks. It includes operators, graph compilation, and runtime management. Huawei also offers ModelArts, a cloud-based platform for model training and deployment, which integrates with Pangu and Ascend. ModelArts supports Data Augmentation, Model Pruning, and Curriculum Learning to improve model efficiency.
Huawei's AI software emphasizes compatibility with industry standards, including onnx and TensorFlow, to ease migration from other platforms. However, due to US export controls, Huawei has developed its own ecosystem to reduce dependence on foreign technology. This includes the openeuler operating system for servers and the harmonyos operating system for devices, which incorporate AI capabilities.
Cloud and Enterprise AI
Huawei Cloud, the company's cloud computing arm, offers AI services powered by Ascend and Pangu. These include pre-trained models for Natural language processing, computer vision, and Speech recognition, as well as custom model training. Huawei Cloud competes with Alibaba Cloud, Amazon Web Services, and Microsoft Azure in the Chinese market, and has expanded to international regions.
In enterprise settings, Huawei AI is used for Predictive analytics, anomaly-detection, and Computer vision applications. For example, in manufacturing, Ascend-based systems inspect product quality; in energy, Pangu models forecast wind power output; in healthcare, AI assists in medical imaging analysis. Huawei has partnered with Intel and AMD for CPU-based servers, but its AI accelerators are increasingly preferred in China due to government procurement policies favoring domestic technology.
Huawei also provides AI solutions for autonomous-driving through its mdc (Mobile Data Center) platform, which uses Ascend chips for real-time processing in vehicles. This aligns with Huawei's broader automotive business, which includes electric vehicle autonomous driving systems.
Research and Development
Huawei AI research is conducted at multiple centers, including the Huawei Noah's Ark Lab, established in 2012, and the Huawei Research Center in Shenzhen. The company invests heavily in R&D, with over 100,000 employees in research roles as of 2023. Huawei has published papers in top AI conferences like NeurIPS, ICML, and CVPR, contributing to areas such as Transformer (architecture) architectures, Multi-Head Attention, and Positional Encoding.
Huawei collaborates with universities, including University of Toronto and University of Oxford, on AI research. It also participates in open-source projects, such as mindspore and openharmony. Despite US sanctions, Huawei continues to innovate, with patents in AI chip design and model optimization. In 2024, Huawei filed over 5,000 AI-related patents, ranking among the top global applicants.
The company's AI research has led to developments in Residual Network (ResNet) architectures, Batch Normalization, and Layer Normalization techniques, which are integrated into its frameworks. Huawei also explores Reinforcement learning and Reinforcement Learning from AI Feedback (RLAIF) (reinforcement learning from AI feedback) to improve model alignment.
Challenges and Sanctions
Huawei AI faces significant challenges due to US sanctions imposed in 2019, which restricted access to American technology, including NVIDIA GPUs and TSMC advanced fabrication. This has forced Huawei to rely on its own Ascend chips and domestic foundries like smic, which produce less advanced nodes. As a result, Ascend chips lag behind NVIDIA's latest offerings in raw performance, but Huawei compensates with software optimizations and system-level integration.
The sanctions also limited Huawei's ability to use Google Cloud and other US cloud services, leading to the development of its own AI cloud. Additionally, Huawei has been banned from selling equipment in several countries, including the United States, Australia, and Japan, due to national security concerns. These restrictions have impacted its global AI business, but Huawei has pivoted to the Chinese market, where it dominates.
Despite these hurdles, Huawei AI has grown, with Ascend shipments increasing in 2024. The company has also faced accusations of intellectual property infringement, though it has settled disputes with firms like cisco. Huawei argues that its technology is safe and has not shown evidence of espionage, but experts note that Chinese laws can compel cooperation with state intelligence.
Future Directions
Looking ahead, Huawei AI aims to expand its ecosystem and achieve greater self-reliance. The company plans to release more powerful Ascend chips, such as the Ascend 920, and improve Pangu models with larger context windows and multimodal capabilities. Huawei is also investing in quantum-computing and Edge AI to diversify its portfolio.
In 2025, Huawei announced a partnership with Alibaba DAMO Academy to advance AI research, and it continues to explore applications in Robotics and smart-manufacturing. The company is also working on Model Pruning and Quantization to make AI more efficient for deployment on Arm Holdings-based devices. As China pushes for technological independence, Huawei AI is positioned to play a central role in the country's AI infrastructure.
Huawei's AI division is also expanding internationally, targeting markets in Southeast Asia, Africa, and Latin America, where it can offer cost-effective AI solutions. However, geopolitical tensions may limit its growth in Western markets. Despite these challenges, Huawei AI remains a key player in the global AI landscape, with a focus on innovation and resilience.