What is GLM-Image?
GLM-Image is Z.ai's image model for pictures that carry a lot of writing, such as posters, slide graphics and labelled science diagrams. A 9B language model plans the layout and meaning, then a 7B diffusion decoder with a glyph encoder draws the detail and the letters. It is still Z.ai's flagship image model, and no newer version has replaced it.
Z.ai says its general image quality is on par with mainstream diffusion models; its edge is text. It posted the best word accuracy in Z.ai's multi-region text test, closed models included, but trails the smaller Z-Image on the DPG-Bench and OneIG general image tests. The weights also edit and restyle images; the hosted API only turns text into images.
Who is GLM-Image best for?
GLM-Image is best for posters, slides, diagrams and ads where the words must be spelled right, for teams that host their own models, and for developers who want a low-cost image API.
Posters and ads with text
Z.ai built it to spell words correctly inside images, and it scored the best word accuracy of any model in Z.ai's multi-region text test.
Diagrams and explainers
Labelled science illustrations, process diagrams and slide graphics are among the jobs Z.ai tuned it for.
Teams that host their own models
MIT-licensed weights run with diffusers or SGLang on your own GPUs: more than 80GB of memory at full speed, or about 23GB with slower CPU offloading.
Edits in the same model
The open weights handle text-to-image, edits, style transfer and keeping a person or product the same across images; the hosted API does text-to-image only.
What's included with GLM-Image?
Text rendering built into the decoder, editing and style transfer in the open weights, a $0.015-per-image API, and support for diffusers and SGLang.
- Text rendering
- A glyph encoder in the decoder sharpens letter strokes; Z.ai asks you to put the exact words in quotation marks.
- Editing in the weights
- Pass one or more input images to edit a scene, change its style or keep a person or product consistent across a set.
- Hosted API
- $0.015 per image, sizes up to 2048 pixels a side in steps of 32, sync or async calls, and a download link that lasts 30 days.
- Self-hosting
- Runs through Hugging Face transformers and diffusers, or as an SGLang server with OpenAI-style generate and edit endpoints.
- Sizes and ratios
- Outputs from 1K to 2K, with preset sizes for ratios such as 1:1, 3:4, 4:3 and 16:9.
How much does GLM-Image cost?
GLM-Image is free to download and run on your own GPUs, and Z.ai's hosted API charges $0.015 per image if you would rather not run it yourself.
- MIT licence for the model, Apache 2.0 for the code
- More than 80GB of GPU memory, or about 23GB with slower CPU offloading
- Weights on Hugging Face and ModelScope
- Text prompts only, sizes up to 2048 pixels a side
- Default hd quality takes about 20 seconds an image
- Returns a download link that expires after 30 days
Self-hosting costs only your own compute. Z.ai lists its API prices in US dollars, and its older CogView-4 image model costs $0.01 per image on the same API. Full pricing on docs.z.ai. We checked these prices on .
How do readers rate GLM-Image?
No one has rated GLM-Image yet. Be the first.
Sign in to rate. One rating per account.
GLM-Image pros and cons
It scored the best word accuracy in Z.ai's multi-region text test and costs little through the API, but self-hosting needs big GPUs and its general photo quality trails smaller open models.
Pros
- It posted the best word accuracy in Z.ai's multi-region text test, 0.9116 on CVTG-2K against 0.8990 for Seedream 4.5, and early testers praised its dense-text examples.
- The model is MIT-licensed, which allows commercial use with no extra terms, and Reddit users singled out the licence as a plus over more restrictive open models.
- The hosted API costs $0.015 per image, so you can use it without owning an 80GB GPU.
- One set of weights handles generation, editing, style transfer and consistent subjects, which launch users welcomed.
Cons
- Heavy to run: Z.ai asks for more than 80GB of GPU memory, a 1024-pixel image took about 64 seconds on an H100 in its own test, and many launch users balked at that requirement.
- General photos lag behind: Z.ai's own tables put it below the smaller Z-Image on the DPG-Bench and OneIG image tests, and several early testers found its photos and anatomy, hands and feet especially, weaker than Z-Image-Turbo's.
- The hosted API takes text prompts only, so editing and reference images need your own GPUs.
Is GLM-Image open source?
Yes. GLM-Image's code is public under the MIT licence, so you can run it on your own server.
On GitHub
1,079stars
- Last commit
- 20 Mar
- Repository age
- 10 months
- Licence
- MIT
- Self-hosted
- Yes
- Repository
- zai-org/GLM-Image
What are the best GLM-Image alternatives?
GLM-Image is an open source alternative to Hunyuan Image 3.0, FLUX 3 Image Klein and Qwen-Image-3.0. Z-Image and ERNIE-Image do the same job and are open source too.
Hunyuan Image 3.0FreeTencent's open text-to-image and editing model.
FLUX 3 Image KleinFreeBlack Forest Labs' 5.6B open-weights model for image edits, text-to-image and box layouts on one GPU.
Qwen-Image-3.0FreemiumChoose it when you want generation and editing through one hosted API, at about $0.025 per image plus $0.00275 per input image; Z.ai's GLM-Image API takes text only.
Z-ImageOpen SourceChoose it when you run models on a consumer GPU: its Turbo version fits in 16GB of VRAM under Apache 2.0, though it trails GLM-Image on text accuracy in Z.ai's tables.
ERNIE-ImageOpen SourceBaidu's 8B open text-to-image model under Apache 2.0.
Ideogram 4.0FreeChoose it when typography is the whole job and you also want a ready-made app and hosted API; its open weights are for non-commercial use only, unlike GLM-Image's MIT licence.
Questions about GLM-Image
What is GLM-Image?
GLM-Image is Z.ai's open-weight image model for posters, slides and diagrams where the words must be right, free to self-host or $0.015 an image on Z.ai's API.
Who is GLM-Image for?
GLM-Image is best for posters, slides, diagrams and ads where the words must be spelled right, for teams that host their own models, and for developers who want a low-cost image API.
Is GLM-Image free?
Yes. GLM-Image's weights are free to download from Hugging Face or ModelScope and run on your own hardware under the MIT licence. Z.ai's hosted API is the paid option, at $0.015 per image.
How much does GLM-Image cost?
GLM-Image's paid plans are Z.ai API at $0.015 per image.
Is GLM-Image open source?
Yes. GLM-Image's code is public under the MIT licence, so you can run it on your own server.
What are the best GLM-Image alternatives?
The closest alternatives on Recaply are Hunyuan Image 3.0, FLUX 3 Image Klein, Qwen-Image-3.0, Z-Image, ERNIE-Image and Ideogram 4.0.
Where can I try GLM-Image?
On its official website, github.com.
What hardware does GLM-Image need?
Z.ai says GLM-Image needs one GPU with more than 80GB of memory, or several GPUs, and about 23GB with CPU offloading at slower speed. On one H100, a 1024×1024 image took about 64 seconds and a 2048×2048 image about 253 seconds.
Can I use GLM-Image commercially?
Yes. Z.ai releases the GLM-Image model under the MIT licence, and its code and two bundled components under Apache 2.0. Both licences allow commercial use, including for images you sell or use in ads.
Can the GLM-Image API edit images?
No. Z.ai's hosted API takes a text prompt and returns one image. Editing, style transfer and reference images work only with the open weights, run through diffusers or an SGLang server on your own GPUs.
GLM-Image vs Z-Image: which should I use?
Use GLM-Image when an image must carry a lot of correct text, and Z-Image for general photos on smaller hardware. In Z.ai's own tables GLM-Image scores higher on text accuracy, while Z-Image scores higher on the DPG-Bench and OneIG image tests and its Turbo version fits in 16GB of VRAM.
Our verdict
GLM-Image is worth it when the words in an image have to be right and you want open weights or a cheap API.
It scored the best word accuracy in Z.ai's multi-region text test, closed models included, it carries an MIT licence, and the API costs $0.015 per image.
- Skip it for
- General photos and portraits, where smaller open models such as Z-Image-Turbo look better and run on a home GPU.
- Check first
- If you need edits or reference images, make sure you can host the weights, since the API takes text only.
GPT Image 2
Ideogram 4.5
Seedream 4.5
MAI-Image-2.7
Grok Imagine Image 2.0