zai-org/GLM-4: the GLM-4-0414 and GLM-Z1 model repository, and the GLM-4-9B docs it kept

The README now describes GLM-4-0414, and GLM-4-9B moved to README_20240605.md

The repository has carried two generations of documentation. The current README.md is titled GLM-4-0414 Model Series and opens with the April 2025 models. The earlier GLM-4-9B series, released on 2024/06/05, is documented in README_20240605.md, with a Chinese counterpart in README_zh_240605.md. The current Chinese README is README_zh.md.

For a reader this split matters. If you arrive from a tutorial written for GLM-4-9B, the front page will not match it, and the instructions you need are in the dated file. The project updates list ties each file to its release: 2024/06/05 for GLM-4-9B, 2024/06/18 for the technical report at arXiv 2406.12793, 2025/04/14 for GLM-4-32B-0414, and 2025/07/02 for GLM-4.1V-9B-Thinking.

Four 0414 checkpoints split by how long they think

The 0414 series is four models with different jobs. GLM-4-32B-0414 is the 32 billion parameter chat model. GLM-Z1-32B-0414 is a reasoning model built on top of it through cold start, extended reinforcement learning, and further training on mathematics, code and logic, plus general reinforcement learning based on pairwise ranking feedback. GLM-Z1-Rumination-32B-0414 is positioned against OpenAI's Deep Research: it thinks longer on open-ended problems, is trained with end-to-end reinforcement learning graded by ground truth answers or rubrics, and can call search tools during its thinking. GLM-Z1-9B-0414 applies the same techniques to a 9B model aimed at resource-constrained deployment.

Picking one is a choice about latency and task shape. The rumination model is built for research-style writing, not short chat turns.

GLM-4-32B-Base-0414 was trained for agent tasks, not just dialogue

The base model, GLM-4-32B-Base-0414, was pre-trained on 15T of high-quality data that includes a substantial share of reasoning-type synthetic data. Post-training added human preference alignment for dialogue, then rejection sampling and reinforcement learning aimed at instruction following, engineering code and function calling. The stated goal is to strengthen the atomic capabilities agent tasks rely on.

The areas zai-org names for GLM-4-32B-0414 follow from that: engineering code, Artifact generation, function calling, search-based Q&A and report generation. The showcase section backs this with generated output for prompts such as a Python program showing a ball bouncing inside a spinning hexagon under gravity and friction, an HTML version of the same scene, a function plotting drawing board, and a Tailwind CSS mobile UI for a machine learning platform.

The GPT-4o and DeepSeek-V3 comparisons are zai-org's own claims

The model introduction states that GLM-4-32B-0414 performs comparably to OpenAI's GPT series and DeepSeek's V3/R1 series, and that on several benchmarks, such as code generation or specific Q&A tasks, GLM-4-32B-Base-0414 matches larger models like GPT-4o and DeepSeek-V3-0324 (671B). It also calls GLM-Z1-9B-0414 top-ranked among open-source models of its size.

These are vendor statements in a launch README. The wording itself is hedged to several benchmarks and specific tasks, which means the comparison does not cover every workload. If you are choosing between GLM-4-32B-0414 and a larger model for production, the consequence is simple: you have to run your own evaluation on your own prompts, because the README's claim tells you where zai-org found parity, not where you will.

Weights live on Hugging Face; this repository holds inference/, finetune/ and demo/

The model files are not in the Git repository. The project updates link to Hugging Face collections for GLM-4-32B-0414 and GLM-4.1V-9B-Thinking, and those collections are where you download weights. The repository itself carries the code around them: inference/ for running the models, finetune/ for adapting them, and demo/ with two examples, demo/composite_demo/ and demo/intel_device_demo/. resources/ holds supporting files such as WECHAT.md, and a .pre-commit-config.yaml sits at the root for contributors.

The Apache-2.0 LICENSE in the repository covers this code. Weights published on Hugging Face carry whatever license their model pages state, so check that page for the exact checkpoint you plan to ship rather than assuming the repository license applies to it.

Newer GLM releases do not live here

The repository is named GLM-4, but not every model with GLM-4 in its name is developed in it. The 2025/07/02 update announces GLM-4.1V-9B-Thinking, a vision language model, and sends readers to a separate repository, zai-org/GLM-4.1V-Thinking. The project updates list stops there: its four entries cover GLM-4-9B, the technical report, GLM-4-32B-0414 and the GLM-4.1V pointer.

The consequence for someone searching for later GLM versions is that this repository will not have their inference code, fine-tuning scripts or model cards. Treat zai-org/GLM-4 as the home of the 0414 and 9B series, and follow the links in the updates list for anything newer.

Z.ai for free trials, bigmodel.cn for paid service, and no releases to pin

The README splits access three ways. The open-source models can be tried for free at https://chat.z.ai. Commercial GLM model services are sold through https://bigmodel.cn. Running the models yourself means Hugging Face weights plus the code in this repository. Community channels are a Discord server, an X account at Zai_org, and a WeChat group documented in resources/WECHAT.md.

The repository has no GitHub releases, and the last push was on 2026-08-05. Without tags, the main branch is the only version, so a change to inference/ or finetune/ reaches you the next time you pull. Record the commit hash you deploy from, and diff against it before updating.

Editorial conclusion

zai-org/GLM-4 is the right starting point if you want to run or fine-tune GLM-4-32B-0414, the GLM-Z1 reasoning models or the older GLM-4-9B series, using the inference/, finetune/ and demo/ code alongside weights from Hugging Face. It is the wrong place to look for GLM-4.1V-Thinking, which has its own repository, or for any GLM version not listed in its project updates. Before committing, read the license on the Hugging Face model page you plan to deploy, pin a commit since there are no releases (last push 2026-08-05), and test the GPT and DeepSeek comparisons on your own tasks instead of taking them as given.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. README
4. zai-org/GLM-4 on GitHub