【#Tech24H】Zhipu, in collaboration with Huawei, has open-sourced the new-generation image generation model GLM-Image, announcing the completion of the entire process from data to training based on Huawei’s Ascend Atlas 800T A2 devices and the MindSpore AI framework. It is reported that the model soared to the top spot globally on the popular AI open-source community Hugging Face’s rankings within less than 24 hours of being open-sourced. Unlike the commonly used LDM approach in open-source projects, GLM-Image adopts an innovative hybrid architecture of “autoregression + diffusion decoder”, balancing global instruction understanding with local detail depiction. It overcomes the challenges of generating knowledge-intensive scenarios such as posters, PPTs, and educational images, and excels particularly in Chinese character generation tasks, marking a significant step toward exploring cognitive generative models represented by “knowledge + reasoning”, such as Nano Banana Pro.