On November 15th, Huawei Cloud's MaaS platform released the open-source Panghu large model, openPangu-2.1-Pro. The updates focus on three main areas: Agent development, AI coding, and inference optimization.
It utilized over 100,000 specialized datasets to optimize the triggering of skills and the execution of tool calls, with a focus on two key aspects of Huawei’s ecosystem: development using HarmonyOS ArkTS and the generation of Ascend computing operators.
These two scenarios clearly illustrate its purpose. ArkTS is the development language for HarmonyOS applications. General-purpose large models often lack sufficient training data, making it difficult to use them effectively in programming; as for the Ascend operators, they are typically used only by those who specialize in inference optimization on Huawei chips. In these contexts, general-purpose models have limitations, while Huawei’s own models perfectly fill these gaps.
Price: When the input value is 0.8 yuan per million tokens, the output price is 3.2 yuan; however, for cached inputs, the cost is only 0.016 yuan, which is one-fiftieth of the regular input price.
It has already been integrated with Huawei's own solutions: AgentArts, CodeArts, and OfficeAce.
For most developers, it will not replace the primary models they are currently using; however, if you are developing applications for HarmonyOS or plan to deploy your models on Ascend platforms, this price is certainly worth considering.