Huawei announced on July 31, 2026 the open-sourcing of the openPangu-2.0-Pro model with 505 billion parameters, with model weights, basic inference code, and technical reports simultaneously published on open-source platforms.
The model was trained on Huawei's Ascend NPU and employs a sparse Mixture-of-Experts (MoE) architecture, with total parameters of 505 billion but only approximately 180 billion parameters activated during each inference. It supports 512K context length and was trained on approximately 34T tokens of data.
This follows Huawei's earlier release on June 30 of the 92-billion-parameter lightweight version, openPangu-2.0-Flash. Huawei Executive Director Richard Yu had previewed the open-source plan at the HDC conference in June, stating that computing power largely supports other domestic enterprises' needs.
The open-sourcing of such a large-scale model trained on domestic chips demonstrates China's growing capability in autonomous AI computing infrastructure, reducing dependence on foreign technology and strengthening the domestic AI ecosystem.
Source: Huawei Official, July 31, 2026
