On January 14, 2026, Zhipu AI, in partnership with Huawei, officially launched the open-source version of its next-generation multimodal image generation model, GLM-Image. Leveraging China's indigenous Ascend Atlas 800T A2 hardware and the MindSpore framework, this model underwent its entire development lifecycle—from data collection to training—on a fully domestic computing infrastructure, making it the first open-source SOTA (State-of-the-Art) model to achieve this milestone using a full-stack, domestically produced computing power base.

GLM-Image employs a hybrid architecture that integrates an "autoregressive + diffusion decoder" approach, seamlessly blending image generation capabilities with language models. This design empowers it to excel in complex image-text tasks, including scientific illustrations and commercial poster creation. When utilized via API calls, generating a single image costs a mere 0.1 yuan, with a speed-optimized version set to be released shortly.