Yu Chengdong Delivers on His Promise, Huawei Pangu openPangu-2.0-Pro Model and Technical Report Open-Sourced

Huawei has officially announced the open-sourcing of the Pangu openPangu-2.0-Pro model and its technical report, marking a further layout in the field of open-source large models. The public release of the model weights and basic inference code will allow developers to directly access and use this large-scale language model.

openPangu-2.0-Pro is a large-scale Mixture of Experts (MoE) language model trained on Ascend NPUs, with a parameter scale reaching approximately 505 billion, and approximately 1.8 billion activated parameters per token. The model supports a 512k context length, with a total training data volume of approximately 34T tokens, demonstrating strong long-text processing capabilities.

One of the core purposes of this open-sourcing is to provide best practice references for the industry through Ascend native training and inference technology. The open-source Pangu brand is committed to helping developers better utilize Ascend computing power, solving adaptation issues for domestic AI chips during model training and inference processes.

In the construction of the domestic AI computing power ecosystem, providing verified open-source models is crucial. Huawei's move helps lower the threshold for developers using Ascend computing power, accelerates the landing of AI applications in different scenarios, and also enhances the influence of domestic large models in the open-source community.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.