At the highly anticipated Huawei Developer Conference (HDC) 2025, Huawei’s cloud computing sector achieved a significant technological leap. Mr. Zhang Ping’an, Executive Director and CEO of Cloud Computing at Huawei, officially launched the Pangu Large Model 5.5. This update covers five key domains: Natural Language Understanding (NLP), Computer Vision (CV), Prediction, Multimodal, and Scientific Computing, injecting new development momentum and technological value into various industries.
The release of Pangu Large Model 5.5 marks remarkable progress for Huawei in the field of AI foundation large models. During the conference, Mr. Zhang particularly emphasized that Pangu Large Model 5.5 was trained entirely within Huawei Ascend Cloud’s powerful full-stack hardware-software collaborative environment. This achievement not only highlights Huawei’s leading edge in hardware-software integration but also further consolidates the Ascend architecture’s global dominance in large model training and inference. Training based on Ascend enables the model to complete learning and iteration more efficiently and with lower energy consumption.
Specifically, in Natural Language Processing (NLP), Pangu Large Model 5.5 introduces the innovative 718B Deep Thinking Model. As introduced, this is a large model adopting a Mixture of Experts (MoE) architecture, trained by 256 experts. The model demonstrates excellent performance improvements in knowledge reasoning, tool invocation, mathematical calculations, and other aspects, reaching industry-leading levels. With the support of innovative algorithms such as universal computation masking, global dynamic balancing, and grouped mixed experts (MoGE), Pangu Large Model 5.5 achieves efficient training and inference on the Ascend platform, with both MFU (Model Floating Point Utilization) training and single-card inference throughput reaching industry-leading levels. This means that under the same hardware conditions, Pangu 5.5 can process NLP tasks faster and more efficiently.
To further enhance user experience, Pangu Large Model 5.5 has been comprehensively upgraded in efficient long-sequence processing, low-hallucination control, fast-slow thinking integration, and Agent technology. The adaptive fast-slow thinking integration technology is particularly noteworthy. By constructing a difficulty-aware dataset and a two-stage progressive training process, this technology allows the model to flexibly switch thinking modes based on the complexity of problems—responding quickly to simple questions and delving deep into complex ones. This intelligent adjustment not only improves the model’s overall reasoning efficiency by up to 8 times, as claimed, but also makes it more flexible and adaptive in practical applications.
The Pangu deep research tool DeepDiver also stands out. Through key technologies such as long-chain puzzle synthesis and progressive rewards, DeepDiver demonstrates extremely high efficiency in application scenarios like web search and common-sense Q&A. Official data shows that DeepDiver can complete over 10 complex Q&As and generate professional research reports exceeding 10,000 words within just 5 minutes, significantly improving work efficiency. This undoubtedly serves as a powerful auxiliary tool for information gathering, analysis, and report writing.
Beyond the breakthroughs in NLP, Mr. Zhang also detailed the technological upgrades of the Pangu Large Model in basic models for Computer Vision (CV), Prediction, Multimodal, and Scientific Computing. These upgrades not only significantly enhance model performance and accuracy but also foster abundant innovative applications and on-site practices in agriculture, industry, scientific research, and other fields. The Pangu Large Model is gradually becoming a leader in industrial AI, providing strong technical support for solving real-world problems across industries. For example, in agriculture, it can be applied to precision farming, offering intelligent planting suggestions based on crop growth and environmental data; in industry, it can be used for fault diagnosis and predictive maintenance, reducing equipment downtime and improving production efficiency. As technology continues to advance and application scenarios expand, the Pangu Large Model is poised to deliver greater value in more domains.
Chinese 
English