Huawei Open Sources 505 Billion Parameter Pangu Large Model

On July 31, 2026, Huawei announced the open sourcing of its Pangu large model with 505 billion parameters, named openPangu-2.0-Pro. The model weights, basic inference code, and technical reports were simultaneously released on open source platforms. Based on Ascend NPU training, the model adopts a sparse MoE architecture with approximately 18 billion activated parameters per inference, supporting 512K context length, and trained on approximately 34 trillion tokens. This follows Huawei's June release of the 92 billion parameter lightweight version openPangu-2.0-Flash.