← All models

CogView4-6B

Published Publisher Zhipu AI

Model positioning

CogView4-6B is a text-to-image model released by Zhipu. It uses the approximately 6-billion-parameter CogView-4 Diffusers version, accepts text prompts, and generates images. The official team announced the adaptation and open-sourcing of this version on March 4, 2025. (github.com)

Capabilities and uses

  • Supports Chinese and English prompts and can render Chinese text in generated images; it can be used for illustrations, creative visual assets, and images containing text.
  • Uses GLM-4-9B as the text encoder; the official open-source model documentation lists a maximum prompt length of 1024 tokens.
  • Supports inference through Diffusers, and the official team also provides CogKit fine-tuning and inference tools. (github.com)

Limitations and deployment

  • The open-source model requires both image width and height to be between 512 and 2048 pixels and multiples of 32; the total pixel count must not exceed 2²¹.
  • The official model card lists BF16 and FP32 inference precision; FP16 is not supported because it may overflow and produce entirely black images.
  • Hardware requirements depend on resolution and offloading configuration. In tests using BF16, a batch size of 4, and 1024×1024 resolution, the official team recorded approximately 35 GB of GPU memory without offloading, approximately 20 GB with CPU offloading, and approximately 13 GB when also using a 4-bit text encoder. The official team recommends at least 32 GB of system memory.
  • Text in generated images is not guaranteed to be completely accurate, and actual results should be checked by a person. The official model card states that the weights use the Apache 2.0 license; users should still consult the specific license terms. (huggingface.co)

Pricing

The following are hosted service prices published by the official platforms under the CogView-4 name. The pricing pages do not confirm that their deployed versions are identical to the open-source 6B weights described here, so these prices should not be treated as the operating cost of the self-hosted model. Date checked: November 19, 2026.

The mainland China and international prices apply to their respective platforms and should not be regarded as the same plan converted at an exchange rate. Hardware, deployment, and other costs for self-hosted weights depend on the actual environment.

Official sources