ChatGLM

ChatGLM is a family of large language models developed jointly by Zhipu AI (operating as Z.ai) and the Knowledge Engineering Group at Tsinghua University in Beijing. The ChatGLM series is distinguished by its strong bilingual Chinese-English capability, open-weight releases under permissive licensing, and a research lineage rooted in the General Language Model (GLM) pretraining framework developed at Tsinghua.

Background

Zhipu AI was founded in 2019 as a spin-off from Tsinghua University's AI research group, with a mission to advance large-scale language model research and commercialise the resulting models. The GLM pretraining framework, introduced by the Tsinghua team, proposed an autoregressive blank infilling objective as an alternative to BERT's masked language modelling and GPT's left-to-right autoregressive prediction, achieving strong performance on both NLU and NLG tasks with a single unified model.

The first public ChatGLM release — ChatGLM-6B — was made available in March 2023. The 6-billion parameter model, optimised for dialogue and deployable on consumer-grade hardware, attracted widespread adoption among Chinese developers, researchers, and enterprises. Within weeks of its release, ChatGLM-6B became one of the most forked and downloaded Chinese-language LLMs on Hugging Face and GitHub.

GLM-4 Series (2024-2025)

The GLM-4 generation marked a significant maturation of the ChatGLM series. GLM-4 All Tools was specifically aligned to understand user intent and could access online information via a web browser, use the Python interpreter for mathematical computation, leverage a text-to-image model for image generation, and call user-defined functions — establishing ChatGLM as a competitive agentic AI platform.

In 2025, Zhipu AI accelerated its release cadence substantially.

GLM-4.5 (July 2025) features 355 billion total parameters with 32 billion active parameters, using a Mixture-of-Experts architecture that enables efficient inference. A lighter variant, GLM-4.5 Air (106B total, 12B active), targets cost-sensitive deployment. Both models support a "thinking" mode for complex multi-step reasoning and a "non-thinking" mode for fast conversational responses.

GLM-4.5V (August 2025) extended the series to vision-language capabilities, with 106 billion parameters enabling multimodal input understanding.

GLM-4.6 (September 2025), a frontier-scale model at 355 billion parameters, was notable for being trained using China's domestic AI chips, reducing dependence on NVIDIA hardware. It is released under the MIT License, making it one of the few frontier-scale open-weight models that enterprises can self-host, deeply customise, and deploy without API lock-in or usage restrictions.

GLM-4.7 (December 2025), the most recent major release, is an open-weight model emphasising deep mathematical reasoning, multi-file software engineering tasks, and stable agentic orchestration. It supports a 200,000-token context window with a maximum output of 128,000 tokens.

Enterprise and Developer Ecosystem

ChatGLM's consistent open-weight releases have fostered a substantial developer ecosystem. Fine-tuned variants have been produced for medical question answering, legal document analysis, financial report generation, customer service, and code generation, particularly targeting Chinese-language enterprise applications that require domain-specific performance.

Zhipu AI operates Zhipu Qingyan as its commercial consumer product, offering GLM-4-based conversational AI accessible via web and mobile. The company's enterprise platform provides API access, fine-tuning services, and private deployment options for corporate clients.

ChatGLM's open-weight nature makes it particularly relevant to Malaysia's research universities and technically sophisticated enterprises that require on-premises AI deployment for data privacy or compliance reasons. Universiti Malaya, Universiti Teknologi Malaysia, and UTAR (Universiti Tunku Abdul Rahman — a Chinese-medium institution with strong links to Malaysia's Chinese community) have researchers experimenting with ChatGLM fine-tuning for Mandarin-Malay NLP tasks.

For Malaysian enterprises with significant Chinese-language document processing requirements — including legal firms handling contracts with mainland Chinese counterparties, import-export businesses in Johor's trade corridor with Singapore, and manufacturing companies in Penang's electronics cluster dealing with Chinese component suppliers — ChatGLM's strong Chinese comprehension and generation capability is a practical advantage over English-primary models.

The model's deployability on modest hardware (ChatGLM-6B variants can run on consumer GPUs) is significant for Malaysian SMEs that lack the budget for cloud API costs at scale. The GLM-4.6's MIT licensing enables Malaysian software vendors to build and redistribute commercial products on top of the model without revenue-sharing obligations to Zhipu AI.

MDEC's developer programmes and the Malaysia Digital Economy Blueprint's emphasis on building local AI capabilities create an ecosystem where open-source Chinese LLMs like ChatGLM complement Western open-weight models (Llama, Mistral), giving Malaysian developers a broader palette of foundation models to work with, particularly for multilingual Southeast Asian language applications combining Chinese, Malay, and English.

Malaysia's significant Chinese-medium education sector, spanning over 60 independent Chinese high schools and numerous Chinese primary schools, represents an underserved market for Mandarin-capable AI tutoring tools — a use case where ChatGLM's bilingual strength is directly applicable.

See Also

References

1. Du, Z., et al. (2022). GLM: General language model pretraining with autoregressive blank infilling. ACL 2022. Tsinghua University.
2. Zeng, A., et al. (2023). GLM-130B: An open bilingual pre-trained model. ICLR 2023.
3. InfoQ. (2025). GLM-4.5 launches with strong reasoning, coding, and agentic capabilities. infoq.com.
4. Wikipedia. (2025). Z.ai. en.wikipedia.org.
5. GitHub. (2025). GLM-4-0414 model series. github.com/zai-org/GLM-4.