GLM-V ships example code and download links for seven checkpoints, not the model itself
GLM-4.6V/4.5V/4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning
At a glance
- What is it?
- zai-org/GLM-V bundles download links, runnable examples, a training reward system and Skills for the GLM-4.6V, GLM-4.5V and GLM-4.1V vision-language line, but it holds no model implementation, and its own news log dated 2026/09/02 says it will no longer be maintained.
- Who is it for?
- Anyone after GLM-4.6V or GLM-4.5V weights can treat this repository as a directory of links plus a handful of examples, and should budget for the checkpoint swap work that two different conversation templates require. Nobody should plan a long-lived dependency on it.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 10 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- zai-org/GLM-V2,406 ★ · +12
| Repository | 2026-09-29 | 2026-11-15 |
|---|---|---|
| zai-org/GLM-V | 2.4k | 2.4k |
12 Stars in the last 47 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The retirement notice and the push record disagree
The news log opens with an entry dated 2026/09/02 that reads, in full: This repository will no longer be maintained. Questions about GLM-5.3-Flash are sent to a separate GLM-5 repository. The push history says something else. The recorded last push is 2026-10-02, about a month after the sentence that ends maintenance, so the retirement notice and the commit record disagree and neither one can be treated as the real status.
The banner above that log drifts the same way. A visitor is first shown links for GLM-5.3-Flash, a model the retirement entry tells them to discuss somewhere else, plus an API platform entry point and a z.ai trial marked as coming soon. The three families named in the repository description, GLM-4.6V, GLM-4.5V and GLM-4.1V-Thinking, appear only further down, inside the download table. Read the banner and the news log together and the repository is advertising a product line it no longer hosts.
The model implementation is not in this repository
No model is implemented here. The algorithm for GLM-4.5V and GLM-4.6V lives in the transformers repository under `glm4v_moe`, and the algorithm for GLM-4.1V-9B-Thinking lives there under `glm4v`. Both are pointers into somebody else's tree, and the README makes no attempt to vendor a copy.
What zai-org/GLM-V keeps is the wrapper layer, and the top level of its tree says what that is made of: `examples/`, `glmv_reward/`, `inference/`, `skills/`, `resources/`, `requirements.txt`, `.pre-commit-config.yaml`, `.gitignore`, `LICENSE`, `README.md` and `README_zh.md`. There is no `src/` directory and no model definition file anywhere in that list.
The practical boundary matters the moment something breaks. A tokenizer entry that does not appear, a chat template that is silently ignored, a kernel that will not build on your accelerator: the fix belongs to the transformers codebase pinned at `transformers>=5.13.0`, not to anything you can patch in this repository. Looking here for the model definition is looking in the wrong place.
One preprocessing path, two conversation templates
The implementation section carries a warning that is easy to skim past: both models share identical multimodal preprocessing, but use different conversation templates, and the reader is told to distinguish them carefully. Shared preprocessing with a swapped template is the worst shape of difference to have, because the image pipeline still looks correct while the text side quietly stops matching what the checkpoint was trained on.
Nothing in this repository enforces the distinction. There is no configuration file recording which template belongs to which checkpoint, and no runtime check that raises when the two are crossed. The GLM-4.5V and GLM-4.6V pair routes through `glm4v_moe` and GLM-4.1V-9B-Thinking routes through `glm4v`, so the model directory name is one usable handle. The Type column is the other, separating Hybrid Reasoning rows from the plain Reasoning and Base rows.
Seven download rows, two mirrors, no releases
The download table has seven rows. GLM-4.6V, GLM-4.6V-FP8, GLM-4.6V-Flash, GLM-4.5V and GLM-4.5V-FP8 all carry the type Hybrid Reasoning. GLM-4.1V-9B-Thinking is typed Reasoning and GLM-4.1V-9B-Base is typed Base. Each row points at two hosts: a Hugging Face page under `zai-org` and a ModelScope page under `ZhipuAI`, so every checkpoint exists twice under two different owner names and the two mirrors are free to drift apart.
Nothing is versioned from this repository. It has no GitHub releases, which leaves no tag stream, no release artifact and nothing to pin. The only quantized format named anywhere is GGUF, and that is attributed elsewhere too, pointing at a ggml-org collection on Hugging Face rather than a build produced here.
requirements.txt carries the linter and the demo alongside torch
requirements.txt is the only dependency declaration in the tree, and it mixes three different kinds of package in one list without saying which is which.
torch>=2.10.0
torchaudio>=2.10.0
torchcodec>=0.10.0
torchvision>=0.25.0
pre-commit>=4.6.0
PyMuPDF>=1.28.0
av>=18.0.0
spaces>=0.51.0
gradio>=6.20.0
transformers>=5.13.0
accelerate>=1.14.0The inference floor is `torch`, `torchvision`, `accelerate` and `transformers`. The demo surface is `gradio` and `spaces`, which belong to the Hugging Face Space referenced in the news log rather than to running a model at all. `pre-commit>=4.6.0` is a lint hook that pairs with the `.pre-commit-config.yaml` sitting at the repository root, so it is contributor tooling installed next to the runtime. `PyMuPDF`, `av` and `torchcodec` cover document and video container handling, and `torchaudio` covers audio.
Every line is a floor with no ceiling. `transformers>=5.13.0` accepts 5.x, 6.x and whatever comes after, so an upstream breaking release slides underneath you with no signal from this file. There is no lock file and no pinned hash anywhere in the tree to hold the set still.
Bounding boxes are normalized by 1000 and the token spec stops there
Grounding is where the output format is most specific, and the numbers are not pixels. The model returns a quadruple `[x1,y1,x2,y2]` holding the top-left and bottom-right corners of the target. Each value is normalized by the image width for x or the image height for y and then scaled by 1000. A consumer therefore has to divide by 1000 and multiply by the real dimension to recover a pixel coordinate. Reading the quadruple as either a pixel count or a plain 0 to 1 fraction puts the box in the wrong place.
The query side accepts free-form descriptions through an `<expr>` placeholder and can be told which output shape to produce, with two sample prompts given: one asking the model to locate an object and return its bounding boxes, one asking it to pinpoint a box directly in the `[[x1,y1,x2,y2], ...]` form.
The response side is framed by two special tokens, `<|begin_of_box|>` and `<|end_of_box|>`. That grounding section stops right there, in the middle of the sentence that introduces them, so how the two tokens behave is not written down here: whether a serving layer strips them, whether they survive into an API response, and what a downstream parser is supposed to do with them.
The build-from-source link points past the listed example directories
The 2025/08/11 news entry says a handcrafted desktop assistant app was open-sourced alongside GLM-4.5V, able to capture screen content through screenshots or screen recordings once connected to the model. Two routes are offered: a Hugging Face Space link for downloading an installer, and a build-from-source link pointing at `examples/vllm-chat-helper/README.md`.
The example directories the tree carries are `AMD_GPU/`, `Ascend_NPU/`, `gui-agent/`, `midscene-ts-demo/`, `midscene-yaml-demo/` and `vlm-helper/`. `vllm-chat-helper` is not among them. The only documented way to build that assistant from source therefore points at a directory the repository layout does not show, while the installer route depends on a Space hosted outside this repository.
That asymmetry matters when you decide what you are actually receiving. The packaged binary is an artifact you cannot inspect from here, and the source recipe does not resolve against the tree as listed. Anyone planning to customize the assistant into their own multimodal tool should confirm that path exists before designing around it.
One runnable command and two vendor specific example trees
Exactly one command appears in the entire README, and it is neither an install nor a serve command. It is offered for the VLM Reward System released on 2025/07/16, the one component with a top-level package directory of its own, `glmv_reward/`, and it starts a demo rather than a service:
python examples/reward_system_demo.pyThe example tree also splits by accelerator. `examples/AMD_GPU/` and `examples/Ascend_NPU/` sit alongside the vendor-neutral names, so the same category of work has hardware-specific entry points for two very different hardware families, with nothing in the tree saying what actually differs inside them.
With no install command anywhere, weights come from the Hugging Face and ModelScope pages in the table, and serving comes from whatever stack you assemble around them. That is the honest shape of this repository: a directory, a handful of runnable examples, and a training reward component.
Editorial conclusion
Anyone after GLM-4.6V or GLM-4.5V weights can treat this repository as a directory of links plus a handful of examples, and should budget for the checkpoint swap work that two different conversation templates require. Nobody should plan a long-lived dependency on it. The news log dated 2026/09/02 says maintenance is over, the last push landed 2026-10-02, and there are no releases here to pin. Verify the `glm4v_moe` and `glm4v` implementations in transformers and read the GLM-5 repository before committing to anything.
Frequently asked questions
What is GLM-V in the zai-org/GLM-V repository?
It is the open-source home of the GLM-4.6V, GLM-4.5V and GLM-4.1V vision-language series from Zhipu AI. The repository holds example code, a VLM reward system, Skills, inference helpers and download links, not the model implementation itself.
Does zai-org/GLM-V contain the code that runs GLM-4.5V and GLM-4.6V?
No. The algorithm for those two lives in the transformers repository under `glm4v_moe`, and GLM-4.1V-9B-Thinking lives there under `glm4v`. The tree here holds `examples/`, `glmv_reward/`, `inference/`, `skills/` and `resources/` instead.
Which GLM-V checkpoints can be downloaded, and from where?
Seven rows are offered: GLM-4.6V, GLM-4.6V-FP8, GLM-4.6V-Flash, GLM-4.5V, GLM-4.5V-FP8, GLM-4.1V-9B-Thinking and GLM-4.1V-9B-Base. Each links a Hugging Face page and a ModelScope page, and the repository publishes no GitHub releases of its own.
How does GLM-V report where an object is located in an image?
As a quadruple `[x1,y1,x2,y2]` holding the top-left and bottom-right corners. Each value is normalized by the image width for x or the height for y and then scaled by 1000, and the answer is framed by the `<|begin_of_box|>` and `<|end_of_box|>` tokens.
Is zai-org/GLM-V still being worked on?
Its news log carries an entry dated 2026/09/02 saying the repository will no longer be maintained, with GLM-5.3-Flash questions directed to the GLM-5 repository. The recorded last push is 2026-10-02, about a month after that entry.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/zai-org-glm-v)Star history chart for your README
[](https://hysenlabs.com/projects/zai-org-glm-v)Was this useful?