[First release of native adaptation Ascend: Huawei officially opens source 92 billion parameter openPangu-2.0-Flash] According to Beating monitoring, Huawei has announced the official open source of the Pangu openPangu-2.0-Flash model and will soon be launched on the Ascend Tribe community of GitCode. The model has a total of 92 billion (92B) parameters and adopts a mixed expert (MoE) architecture. In reality, only 6 billion parameters are activated and it supports 512K ultra long contexts. As a Huawei open-source AI model brand, openPangu is committed to providing the industry with practical reference for Ascend computing power development through Ascend's native training and inference technology. According to the open source plan, the first batch of Flash versions to be launched includes model weights, basic inference code, and training and inference operators; The more powerful openPangu-2.0-Pro weight and inference code will be released in July, and more components will be open sourced in the second half of the year. [Original link]