GLM-Image is a new open-source model designed for high-quality image generation. It combines an auto-regressive module, based on a 9 billion parameter model called GLM-4-9B-0414, with a diffusion decoder inspired by CogView4, which has 7 billion parameters. This hybrid architecture excels in tasks requiring a deep understanding of complex information and precise semantic expression. It performs well in text-to-image generation and supports various image-to-image tasks like editing and style transfer. The model addresses limitations found in traditional diffusion models, particularly in following complex instructions and representing dense knowledge. By using semantic-VQ tokens as its primary tokenization method, GLM-Image ensures better semantic correlation in visual generation. The auto-regressive generator focuses on low-frequency semantic signals, while the diffusion decoder adds high-frequency details to create a polished final image. This separation allows for improved performance in creative tasks requiring intricate knowledge representation. For image editing, the model integrates both semantic-VQ tokens and VAE latent representations from reference images to preserve fine details. Unlike other models that employ full attention mechanisms, GLM-Image uses a block-causal attention approach, reducing computational overhead without sacrificing detail preservation. In the post-training phase, a decoupled reinforcement learning strategy optimizes both the generator and decoder separately, enhancing semantic alignment and visual quality. The training employs GRPO optimization techniques tailored for diffusion models, focusing on aesthetic and semantic consistency in the generated outputs.