Huawei unveils new large-scale AI model aimed at enterprise

Huawei's new large-scale AI model, Pangu 3.0, confirms the Chinese company's focus on enterprise AI. It aims to provide AI that supports complex industry use cases rather than powering chatbots, as Huawei Cloud CEO Zhang Pingan has put it.

"Pangu does not compose poetry, but does things," Zhang said when introducing the new model at a Huawei developers' conference early this month. He described Pangu 3.0, which is already deployed in ten industries, as a large model system designed to meet industry needs.

Named after a mythical Chinese creation figure, Pangu 3.0 consists of three layers: the foundational large language model (LLM), the industry-specific model and a scenario-specific model.

The decoupled architecture means the Pangu models can be adapted to different tasks, allowing customers to load their own datasets and train their own models, Zhang says. He describes it as "a system of pre-trained models that can be quickly adapted to meet scenario-specific needs and address complex challenges across multiple industries."

Zhang cited Pangu's AI weather prediction model, which is claimed to be the first to demonstrate higher precision than traditional numerical weather forecast methods. It had delivered a 10,000-fold improvement in the speed of weather prediction, reducing prediction time to just seconds, and had accurately predicted the trajectories and landing times of typhoons.

AI cloud as well

In the mining sector, another key vertical, Pangu 3.0 made its debut Tuesday for Huawei customer Shandong Energy Group. It enables unsupervised self-learning in 300 coal-mining scenarios, achieving 98% accuracy in tracking the key metric of transport failures.

Huawei began building the Pangu large learning model in 2019 and completed the initial version in April 2021.

In tandem with the new AI model, the company has also released its Ascend AI cloud service, with a single cluster capable of 2,000 petaFLOPS and able to train a multibillion parameter model for 30 days.

Zhang said Huawei had sought to build out its own AI computing power because of the short supply of computing capacity in China. "Now there are hundreds of projects in the queue. The computing power buildout has been slow, the price of GPU is high and the delivery cycle is very long," he noted.

Zhang added that Huawei aims to meet the needs of Chinese industry customers and become "another pole of computing power" in China.

Related posts:

— Robert Clark, Contributing Editor, special to Light Reading