CogView4

Chinese text rendering inside generated images was CogView4-6B's headline feature: at its release in March 2025 it was the first open weights text-to-image model to draw Chinese characters natively, scoring 0.6168 on the team's Chinese text accuracy F1 benchmark against 0.2880 for Kolors, the open competitor it was measured against. The 6B parameter diffusion transformer from Zhipu AI, now Z.ai, pairs with the GLM-4-9B text encoder, accepts prompts up to 1024 tokens in Chinese and English, and generates at arbitrary resolutions from 512 to 2048 pixels per side as long as dimensions are divisible by 32 and the total stays under 2^21 pixels. Usage goes through Hugging Face diffusers via CogView4Pipeline in BF16 or FP32, since FP16 produces black images; memory needs run from roughly 13 GB of VRAM with CPU offload and a 4-bit quantized text encoder up to 39 GB with no offloading at high resolution, with 32 GB of system RAM recommended. Both the code and the model weights are Apache-2.0 licensed, permitting commercial use, and the companion CogKit toolkit from the same team covers fine-tuning.