Huawei has officially open-sourced its openPangu-2.0-Flash model, landing soon on GitCode's Ascend Tribe community.
The model packs 92 billion total parameters using a mixture-of-experts (MoE) architecture, but only activates 6 billion parameters per inference. It supports a 512K-token context window — enough to chew through entire book-length documents. As Huawei's open-source AI brand, openPangu aims to give developers a real-world reference for training and inference on its in-house Ascend processors.
The Flash release includes model weights, basic inference code, and training/inference operators. A beefier openPangu-2.0-Pro version — with its own weights and inference code — is slated for July, with more components trickling out through the rest of the year.