CogView-4

CogView-4 is the image generator Z.ai published in March 2025 and still sells today, a year and a half later, as the cheap alternative to its own flagship. At $0.01 per image it is the least expensive model in the company's entire price list, and a third of the price of GLM-Image, which replaced it at the top of the range in January 2026. Its distinguishing feature at launch was linguistic rather than visual. The text encoder is GLM-4-9B, a bilingual language model, which means the generator reads Chinese instructions directly instead of routing them through an English translation layer — and can render Chinese characters inside the picture. Most competing generators of that period handled Chinese only as a translated afterthought. The technical envelope is unusually flexible for a model of this generation. Width and height are free anywhere between 512 and 2048 pixels, each divisible by 32, with the only hard limit being a total of about two million pixels — so the model is not confined to a fixed set of aspect ratios. The published memory table is honest about the cost: roughly 35 GB of video memory at 1024x1024 with default settings, down to about 13 GB with CPU offloading and a 4-bit text encoder. The card also warns that FP16 is not supported at all, because it overflows and returns completely black images. The weights are on Hugging Face under a full Apache 2.0 licence, with the licence text included in the repository rather than merely declared in the metadata. The safetensors index totals 6.37 billion parameters across the transformer, the text encoder and the VAE, which is where the repository name CogView4-6B comes from.

▸Videos

No videos yet.

▸Reviews