June 20, 2012 - In this afternoon'sHuaweiManaging Director of Huawei during the HDC 2025 keynote at the developer conference,HUAWEI CLOUDComputing CEO Pingan Zhang officially launchesPangu Model 5.5, the five basic models of Natural Language Processing (NLP), Computer Vision (CV), Multimodal, Prediction, and Scientific Computing are fully upgraded.
1AI organizes the main elements as follows:
Pangu Natural Language Processing NLP Grand Model:
The new 718B Deep Thinking Model is a MoE Big Model consisting of 256 experts, which is significantly enhanced to achieve leading capabilities in the areas of knowledge reasoning, tool invocation, and math. The Pangaea Big Model is trained on full-stack hardware and software based on the Rise Cloud.This signifies that it is possible to build world-class models based on the Rise architecture..
Pangu Big Model 5.5 is upgraded with features such as efficient long sequences, low illusion, fast and slow thinking fusion, and Agent to improve user experience. For example, Pangu proposes the technology of adaptive fast and slow thinking in one, which realizes the model to adaptively switch fast and slow thinking according to the difficulty level of the problem by constructing difficulty-aware fast and slow thinking data and two-stage progressive training.Agile responses to simple questions, deep thinking on complex questionsPangu DeepResearch DeepDiver is the first deep research tool in the world to achieve a 8x improvement in overall model inference efficiency. Pangu DeepResearch DeepDiver can achieve high execution efficiency on applications such as web search and general knowledge Q&A through key technologies such as long-chain puzzle synthesis and progressive rewarding.For example, it can complete more than 10 jumps of complex questions and answers in 5 minutes, and generate more than 10,000 words of professional research reports.The company's workforce has been greatly enhanced by the use of the "new" technology and the "new" technology.
Pangaea predicts large models:
Adopting the industry's first triplet transformer unified pre-training architecture, data from different industries, including tabular data of process parameters, time-series data of equipment operation logs, and picture data of product inspection are encoded in unified triplets and efficiently processed and pre-trained within the same framework, which greatly improves the accuracy of the prediction model, and dramatically improves the generalizability of the cross-industry, cross-scene generalizability.
Pangaea Scientific Computing Large Models:
Huawei Cloud continues to expand the integration of the Pangu scientific computing model with more scientific applications. The Shenzhen Meteorological Bureau further upgraded its "Zhiji" model based on Pangu.First AI-assembled forecastThe model can reflect more intuitively the possibility of weather system evolution and reduce the error of a single forecast model. Chongqing Municipal Meteorological Bureau has built the "Tianzhi-12h" meteorological model based on Pangu, which is characterized by strong local precipitation and high precipitation intensity in the Chengdu-Chongqing area, and enhances the ability of intra-day forecasting and early warning of catastrophic weather. Shenzhen Energy uses Pangu to forecast short- and medium-term wind and water power generation capacity, reducing power abandonment on the power generation side and improving the efficiency of energy development.
Pangu computer vision CV macromodel:
Huawei Cloud Releases 30 Billion Parameter Visual Megamodel with New MoE Architecture.This is the largest visual model in the industry todayIt also fully supports multi-dimensional and pan-visual perception, analysis and decision-making such as image, infrared, laser point cloud, spectrum, radar and so on. In addition, Pangu CV Big Model builds the scarce pan-visual fault sample library of oil and gas, transportation, coal mine and other industrial scenarios by generating models across dimensions, which greatly improves the identifiable types and accuracy of business scenarios.
Discordian multimodal macromodeling:
New release of a world model based on the Pangaea multimodal macromodel.It is possible to build the digital physical space needed for intelligent driving and training of embodied intelligent robots., to achieve continuous optimization iteration. For example, in the field of intelligent driving, by inputting the first frame of the driving scene, driving control information and road network data, the Pangu World Model can generate the driving video of each road camera and the point cloud of the LIDAR, which is able to generate a large amount of training data for intelligent driving without relying on high-cost road mining.