Zhipu AI Launches Image Model Trained on Chinese Hardware

The Beijing-based company Zhipu AI released GLM-Image, a state-of-the-art multimodal image generation model trained entirely on Chinese-made hardware. The development follows a decision by the United States Department of Commerce to add the firm to an entity list over alleged ties to the Chinese military, which blocked the company's access to Nvidia H100 and A100 GPUs.

To bypass these Western hardware restrictions, Zhipu AI utilized Huawei Ascend Atlas 800T A2 processors and the MindSpore AI framework to complete the training pipeline. The resulting model employs a hybrid architecture featuring a 9-billion-parameter autoregressive model and a 7-billion-parameter diffusion decoder. According to the company, this achievement proves the feasibility of training high-performance multimodal generative models on a domestically developed full-stack computing platform.

GLM-Image currently ranks first among open-source models on the CVTG-2K benchmark for text placement accuracy. Zhipu AI has made the model weights available on GitHub, Hugging Face, and ModelScope Community, while also providing API access at a cost of 0.1 yuan per image.