GLM-4
Multilingual large language model series by Zhipu AI
GLM-4 is a family of large language models developed by Team GLM at Zhipu AI, announced in June 2024. Pre-trained on ten trillion tokens, it supports long contexts and tool use, aiming to rival top proprietary models in general reasoning and Chinese language tasks.
- Pre-trained on ten trillion tokens
- Supports context windows up to 1 million tokens
- Outperforms GPT-4 on Chinese AlignBench metrics
- Open-weights available for the 9B variant
- Autonomous tool use for web and code tasks
What it is
GLM-4 represents the fourth generation of the ChatGLM family, a series of large language models created by Team GLM. The release includes several variants such as GLM-4, GLM-4-Air, and the open-weight GLM-4-9B. These models are designed for general purpose use, with strong capabilities in reasoning, coding, and multilingual support. The system is built to handle complex tasks by autonomously selecting and using external tools like web browsers or Python interpreters.
How it came about
The development builds on three prior generations of ChatGLM models. Team GLM introduced the series in June 2024 through a technical report. This launch followed the open-sourcing of earlier models like ChatGLM-6B, which had accumulated millions of downloads. The team continued to refine the architecture, incorporating lessons from previous iterations to improve alignment and performance across diverse linguistic and logical tasks.
Inside the model
The models were pre-trained on a dataset of ten trillion tokens, primarily consisting of Chinese and English text, with smaller amounts of data from 24 other languages. Alignment for human preference was achieved through a multi-stage process involving supervised fine-tuning and learning from human feedback. The architecture supports a context window of 128,000 tokens, with specific versions extending this to one million tokens. It integrates capabilities for vision processing and function calling.
What it can do
Evaluations indicate that GLM-4 performs comparably to GPT-4 on benchmarks such as MMLU, GSM8K, and MATH. It shows strong results in instruction following, nearing the performance of GPT-4-Turbo. In Chinese language tasks, it reportedly outperforms GPT-4 on the AlignBench metric. The model demonstrates proficiency in long-context retrieval and can execute code or browse the web to solve problems, matching proprietary counterparts in practical applications.
How to use it
Zhipu AI has made the GLM-4-9B variant available with open weights under an open-source licence. This version supports context lengths of 128K and 1M tokens. Developers can access the model through platforms like Hugging Face. The full GLM-4 series is also accessible via application programming interfaces for commercial and research use, allowing integration into various software environments without local hosting.
Why it matters
The open-sourcing of the 9-billion parameter model has contributed to the broader community, attracting significant download activity. By providing high-performance weights with long-context support, the release enables researchers and developers to build applications requiring extensive document analysis or multilingual support without relying solely on closed proprietary systems.
Questions people ask about GLM-4
What languages does GLM-4 support?
The model is primarily aligned for Chinese and English usage. It also includes support for 24 other languages, including Japanese, Korean, German, French, and Arabic, making it suitable for multilingual applications.
Can GLM-4 use external tools?
Yes, the GLM-4 All Tools model is designed to understand user intent and autonomously decide when to use tools. It can access web browsers, run Python code, generate images, and call user-defined functions to complete complex tasks.
Is GLM-4 open source?
The GLM-4-9B variant is released with open weights under an open-source licence. Larger models in the GLM-4 family are available via API, but their weights are not publicly downloadable.
How long is the context window?
The standard GLM-4-9B model supports a context length of 128,000 tokens. A specialized version of the model is available that extends this capacity to one million tokens, allowing for the processing of very large documents.
Good reading and sources
There is no English Wikipedia article on GLM-4 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2406.12793, huggingface.co/zai-org/glm-4-9b-chat. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.