en.Wedoany.com Reported - On June 30, Huawei officially open-sourced the openPangu-2.0-Flash model with 92 billion parameters. The model weights, basic inference code, and training/inference operators were simultaneously released on the open-source platform. openPangu is Huawei's open-source AI model brand, primarily built around Ascend-native training and inference technologies, providing developers, enterprises, and research institutions with a reference for large model development and deployment. Huawei previously released openPangu 2.0 during HDC 2026, clarifying that the series includes two versions, Pro and Flash, and plans to gradually release related components starting June 30.
openPangu-2.0-Flash has a total parameter count of 92 billion, with 6 billion activated parameters, and supports a 512K context window. The Flash version emphasizes inference efficiency and deployment adaptability, making it suitable for Agent tasks, long-text processing, enterprise knowledge base Q&A, and intelligent application calls within the HarmonyOS ecosystem.
According to Huawei's announced schedule, the openPangu-2.0-Flash model weights, basic inference code, and training/inference operators were made available on the open-source platform on June 30. The openPangu-2.0-Pro model weights and basic inference code will be released in July, with more open-source components to be gradually released in the second half of the year. IT Home reported that openPangu 2.0 plans to gradually open-source seven major components, including model weights, inference code, technical reports, model architecture, as well as newly added pre-training code, post-training code, and training operators. With the expanded scope of open-source components, developers can not only call the model but also observe the engineering pipeline of model training, inference, and operator adaptation.
openPangu 2.0 is divided into two versions: Pro and Flash. Public information shows that openPangu-2.0-Pro has a total parameter count of 505B, with 18B activated parameters; openPangu-2.0-Flash has a total parameter count of 92B, with 6B activated parameters. Both versions support a 512K context window, providing longer input windows for long documents, multi-turn tasks, enterprise knowledge bases, code materials, and Agent workflows. The Flash version has a lighter parameter scale, primarily serving low-latency, high-throughput, and more flexible deployment needs. The Pro version has a larger parameter scale and will release model weights and basic inference code in July.
The focus of this open-source release is on the Ascend-native pipeline. Common challenges when deploying large models on domestic AI hardware include operator adaptation, inference throughput, memory usage, long-context efficiency, and engineering deployment stability. By open-sourcing model weights, inference code, and training/inference operators together, Huawei allows developers to directly reproduce the inference process around the Ascend environment, reducing the debugging cost for enterprises deploying open-source models on domestic computing power. For industry customers in government, finance, manufacturing, transportation, and energy sectors requiring private deployment, whether the model can run stably on local computing platforms often affects application deployment speed more than simple benchmark scores.
Agent applications are a key use case for openPangu 2.0. The 512K context window can accommodate longer task records, tool invocation chains, business rules, and enterprise documents, making it suitable for tasks such as contract review, ticket processing, R&D assistance, customer service knowledge bases, code understanding, and process automation. After the Flash version is released, developers can first build applications around lightweight inference, tool invocation, and industry knowledge bases, then choose the subsequent Pro version based on task complexity. Huawei's Xiaoyi Claw has previously integrated the openPangu 2.0 Pro model to enhance system-level task execution capabilities in HarmonyOS, indicating that openPangu will form closer technical connections with HarmonyOS, Ascend, and enterprise AI applications in the future.
With the official open-source release of openPangu-2.0-Flash, Huawei has created a new connection point between AI models, domestic computing power, and the operating system ecosystem. The open model weights address the "can it be used" issue, while the basic inference code and training/inference operators solve the "how to run it well on Ascend" problem. After subsequent components such as pre-training code and post-training code are released, enterprises and developers can also perform adaptive training on industry data. Key points for future observation include the open-source license terms, component completeness on the GitCode platform, Ascend hardware deployment barriers, actual long-context costs, and the model performance after the Pro version is released in July.
