Zhipu AI GitHub: the open-source repository overview

What lives in the public GitHub organisation

The THUDM organisation on GitHub is the home of inference code, model cards, fine-tuning toolkits, evaluation harnesses, and contribution guides for the GLM and ChatGLM family.

The zhipu ai github public presence is organised under the THUDM account, which takes its name from the Tsinghua University Department of Computer Science research group that co-developed the GLM architecture. The organisation holds more than ten active public repositories spanning the full model lifecycle: from training and evaluation through to inference, fine-tuning, and deployment. Most repositories include English-language documentation alongside Chinese, and the issue tracker is actively monitored by the team.

The flagship repository in the organisation is the ChatGLM inference repository. It contains the reference implementation for loading and running ChatGLM and GLM models using the Hugging Face transformers library, along with example notebooks for single-turn and multi-turn conversations, quantisation guides, and an API server wrapper that exposes the loaded model on a local port with an OpenAI-compatible interface. This last feature makes the ChatGLM repository the fastest path to a local self-hosted GLM server for developers who want to test against the real model without an internet dependency.

Key repositories and what each does

A map of the most useful THUDM repositories for developers, researchers, and fine-tuners.

The ChatGLM series has its own repository per major generation. ChatGLM-6B was the original release that brought the family into global attention; ChatGLM2-6B and ChatGLM3-6B followed with extended context windows and improved instruction-following. The GLM-4 and GLM-4.5 generations moved to a unified repository that covers the full parameter sweep rather than a single 6B target. Each repository follows the same README structure: installation steps, basic usage, quantisation options, and a link to the Hugging Face model page for downloading the weights.

The GLM-4 fine-tuning repository is separate from the inference code. It provides the supervised fine-tuning (SFT) and RLHF pipeline scripts used by the Zhipu AI team, adapted for external use on custom datasets. The repository includes example dataset schemas, a training configuration reference, and guidance on hardware requirements per model size. Community contributors have extended it with LoRA and QLoRA adapters that make fine-tuning practical on single-GPU setups at the 6B and 9B scales.

The evaluation harness repository hosts the benchmark suite the team uses internally. It covers a range of standard benchmarks — MMLU, GSM8K, HumanEval, and several Chinese-language benchmarks not available in the standard English evaluation stacks. Researchers who want to reproduce published GLM benchmark numbers or run comparative evaluations with consistent methodology should start there rather than assembling their own harness from scratch.

Five key THUDM repositories — purpose and update cadence

Repository	Purpose	Cadence
ChatGLM / GLM-4 inference	Reference inference code, local API server, quantisation guides	Updates on model release; patches for community-reported issues between releases
GLM-4 fine-tuning	SFT and RLHF pipeline scripts, LoRA/QLoRA adapters, dataset schemas	Incremental — new adapter support and dataset examples added between model releases
GLM evaluation harness	Benchmark suite covering English and Chinese benchmarks used in published results	Updated when new benchmarks are added to published evaluations
GLM model cards (HF mirror)	Model card content and metadata mirrored from Hugging Face for reference	Per release — each new model ships with an updated card
AgentBench / CogAgent	Agent evaluation benchmarks and multimodal agent research code from the THUDM team	Research releases — updated alongside paper publications

Release cadence and how to track it

The zhipu ai github organisation releases on an aggressive cadence that rewards watching the key repositories rather than checking periodically.

Model releases on the THUDM GitHub typically follow a predictable sequence: a tag on the inference repository, a README update with the new model name and Hugging Face link, and a concurrent push of the weights to Hugging Face. The gap between announcement and usable GitHub code is often under 24 hours; the gap to community GGUF mirrors is typically 24 to 72 hours after that. Watching the ChatGLM and GLM-4 repositories on GitHub — using the Watch button with "Releases only" notifications — is the least-noise way to stay current without checking daily.

Between model releases, the repositories receive regular maintenance: dependency bumps, issue responses, documentation corrections, and occasionally new notebook examples that respond to common community questions. The issue tracker is more active than many comparable research-lab repositories; questions about quantisation, hardware compatibility, and API usage patterns receive responses from maintainers within a few days in most cases.

Contribution patterns

What the THUDM repositories accept from external contributors and where community pull requests tend to land.

External contributions are accepted but scoped. The CONTRIBUTING guides in each repository specify that documentation improvements, example additions, and bug fixes are the most welcome pull request types. Core model architecture code is not maintained externally; the training infrastructure is proprietary and not present in the public repositories at all. The fine-tuning repository is more open to community additions — new adapter types, dataset loading utilities, and evaluation hooks have been merged from contributors who opened well-documented pull requests with test evidence.

Filing an issue before opening a pull request is the recommended path for anything beyond a trivial fix. The maintainers are responsive to issues and will indicate whether a proposed addition aligns with the project roadmap before a contributor invests in a full implementation. Research on open-source collaboration norms from MIT's Open Source Program Office is a useful frame for any team setting internal policies around external open-source contribution. The practical norm at THUDM matches most well-maintained academic lab repositories: issues are the primary feedback channel, and pull requests that reference an existing issue get reviewed faster.