What is GLM-Image?
GLM-Image was released in January 2026 by Z.ai, the company formerly called Zhipu AI. It pairs a 9-billion-parameter autoregressive generator, initialised from GLM-4-9B, that plans the image as about 256 tokens and expands it to 1K to 4K tokens, with a 7-billion-parameter diffusion decoder that has a dedicated glyph encoder for text. One model handles text-to-image and image-to-image work, including editing, style transfer and identity-preserving generation of people and objects.
The main weights are MIT-licensed; the VQ tokenizer and ViT weights come from X-Omni under Apache 2.0. Output runs from about 1K to 2K pixels a side, in dimensions divisible by 32. Z.ai's API charges $0.015 per image. The Hugging Face weights were last updated on 15 January 2026 and the GitHub repository has had no push since March 2026.
How much does GLM-Image cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | MIT, with Apache 2.0 tokenizer and ViT components |
| Z.ai API | $0.015 per image |
Free weights under MIT (with Apache 2.0 components). Z.ai API $0.015 per image.
Is GLM-Image worth it?
GLM-Image is a specialist for text-heavy pictures and a weak general image model. On CVTG-2K, a test of rendering several pieces of text in one image, Z.ai reports word accuracy of 0.9116, ahead of Seedream 4.5 at 0.8990 and Z-Image at 0.8671, and its Chinese LongText-Bench score of 0.979 is the best in the same table.
General quality is where it falls short. On Artificial Analysis's text-to-image board it ranks 89th of 108 at 891 Elo, behind ERNIE Image at 909, Z-Image Turbo at 936 and the open leader Qwen-Image-2.1 at 1045. In Z.ai's own tables its OneIG-EN score of 0.528 ties Z-Image-Turbo and sits below Qwen-Image at 0.539.
It is also slow to run. Z.ai's figures for one H100 show 64 seconds and 37.8GB of memory for a 1024x1024 image, and 253 seconds and 45.1GB at 2048x2048; CPU offloading brings memory down to about 23GB at a further cost in speed. Faster vLLM-Omni and SGLang paths are still listed as being integrated, ten months after release.
The licence and the API are the redeeming parts. MIT on the main weights means no revenue cap and no regional exclusions, and at $0.015 an image Z.ai's hosted version spares most teams the H100 altogether.
Where it earns its place
A restaurant chain's marketing team in Shanghai
Produce bilingual menu boards and promotional posters with exact prices and dish names
Send prompts with every string in quotation marks to Z.ai's API at $0.015 an image, so 2,000 drafts a month cost $30. A designer still proofreads, since 0.9116 word accuracy still leaves errors.
An infrastructure engineer at a publisher with spare H100s
Self-host an image service that clears legal review quickly
Serve the Diffusers pipeline on one H100 at about 64 seconds and 38GB per 1024x1024 image, roughly 56 images an hour per GPU. MIT terms need no revenue or territory checks.
A catalogue team at an online furniture retailer
Swap backgrounds while keeping products and models consistent
Use image-to-image mode with one or more reference pictures; height and width must be set explicitly in multiples of 32, and each 1024x1024 edit takes about 59 seconds on an H100.
Before you commit
- Setup
- Install Transformers and Diffusers from source and load GlmImagePipeline on an 80GB GPU, or enable CPU offloading for about 23GB; the API needs only a Z.ai key.
- Integrations
- Transformers and Diffusers pipelines, an SGLang server with OpenAI-style generation and edit endpoints, and Z.ai's API; Z.ai recommends GLM-4.7 for prompt enhancement.
- Data and privacy
- MIT-licensed weights can run fully offline; the model card publishes no retention terms specific to the hosted image API.
- Cost at scale
- API: 100,000 images a month cost $1,500. Self-hosting: about 56 images an hour per H100 at 1024x1024 with default settings.
- Limits
- Dimensions must be divisible by 32; text to render must be quoted; sampling at temperature 0.9 trades stability for variety; optimised serving is not yet available.
What’s good
- MIT licence with no revenue cap or regional limits
- Highest CVTG-2K word accuracy in Z.ai's comparison
- Text-to-image and editing in one model
- Hosted API at $0.015 per image
What’s not
- 89th of 108 on Artificial Analysis's text-to-image board
- About 64 seconds per 1024x1024 image on an H100
- Needs about 38GB of GPU memory without offloading
- Text to be rendered must be put in quotation marks
Best for
Teams producing Chinese or English posters, menus and infographics who need correct in-image text and an unrestricted licence more than speed or polish.
GLM-Image alternatives
GLM-Image: common questions
- Is GLM-Image worth it?
- GLM-Image gets the words on a poster right more often than anything in Z.ai's own comparison, but it trails on everything else and takes a minute per picture on an H100. We rate it 3.3 out of 5.
- How much does GLM-Image cost?
- Free weights under MIT (with Apache 2.0 components). Z.ai API $0.015 per image.
- What are the downsides of GLM-Image?
- 89th of 108 on Artificial Analysis's text-to-image board; about 64 seconds per 1024x1024 image on an H100; needs about 38GB of GPU memory without offloading; text to be rendered must be put in quotation marks.
- Who is GLM-Image best for?
- Teams producing Chinese or English posters, menus and infographics who need correct in-image text and an unrestricted licence more than speed or polish.
- How much does GLM-Image cost for a team or in production?
- API: 100,000 images a month cost $1,500. Self-hosting: about 56 images an hour per H100 at 1024x1024 with default settings.
- Is GLM-Image safe for company data?
- MIT-licensed weights can run fully offline; the model card publishes no retention terms specific to the hosted image API.
- What are the best GLM-Image alternatives?
- The closest alternatives we have reviewed are ERNIE-Image (and Turbo), Qwen-Image-2.1, MOSS-Image and Ming-Image-0.1-Design.
Answers checked Nov 7, 2026.