GLM-Image

GLM-Image is Z.ai's flagship image generator, published in January 2026. What separates it from most of the field is stated openly in the documentation: it is built for knowledge-heavy pictures — posters, slides, explanatory diagrams — rather than for photorealism or art. Those are the cases where generators traditionally fail, because they demand embedded text that reads correctly, a legible visual hierarchy and consistency across several panels. The architecture is a hybrid the company calls autoregressive plus diffusion decoder: the autoregressive part carries the global understanding of the instruction, the diffusion decoder renders local detail. Z.ai positions this as its entry into the "cognitive generation" approach exemplified by Nano Banana Pro. The documented uses follow from it: commercial posters with precisely placed text, popular-science diagrams with annotations and process arrows, multi-panel e-commerce or comic layouts where the subject has to stay recognisable from frame to frame, and social-media graphics with complex layout. The practical parameters are modest and clearly stated. Text goes in, an image comes out — delivered as a URL that has to be downloaded, not as inline data. Aspect ratios include 1:1, 3:4, 4:3 and 16:9; width and height are free between 512 and 2048 pixels as long as each is a multiple of 32. Generation costs $0.015 per image, half again as much as the older CogView-4 at $0.01. Unusually for an image model, the weights are public: Z.ai released GLM-Image on Hugging Face under an MIT licence, and the safetensors index puts it at 6.93 billion parameters — small enough to run on a single accelerator, which is a large part of the point.

▸Videos

No videos yet.

▸Reviews