Huawei has officially announced that the openPangu-2.0-Flash model weights with 92 billion parameters, basic inference code, and training and inference operators are now officially available on the open-source platform. As the flagship brand of Huawei's open-source AI models, openPangu aims to provide industry references for effectively utilizing Ascend computing power through native training and inference technologies on Ascend. In the context of global large model technology evolving towards long-text processing and low latency, this move is intended to accelerate AI commercial innovation, further build an intelligent infrastructure for the Agent era, and enrich the Ascend developer ecosystem.

At the previous Huawei Developer Conference HDC2026, Huawei officially launched openPangu 2.0. This large model features a 512K ultra-long context processing capability and adopts a differentiated configuration matrix strategy, including Pro and Flash versions to meet different industrial-grade application scenarios. Among them, the 2.0-Flash version, which was released first today, has a total parameter count of 92B and an activated parameter count of 6B, ensuring lightweight operation while maintaining high concurrency inference performance.

According to Huawei's open-source roadmap, related components of openPangu 2.0 will be gradually released starting from June 30th. After the release of the Flash version today, the openPangu-2.0-Pro model weights with a total parameter count of 505B and activated parameter count of 18B, along with the basic inference code, will also be launched on the open-source platform in July. More open-source components are planned to be released throughout the second half of this year.