# ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools

**Authors:** Team GLM,  :, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li  
**arXiv:** [2406.12793](https://arxiv.org/abs/2406.12793)  
**Date:** 2024-06-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2406.12793v1/

## Key points

- Introduction of the ChatGLM family of large language models, focusing on GLM-4 series (GLM-4, GLM-4-Air, and GLM-4-9B)
- Training on vast data in multiple languages with emphasis on Chinese and English
- Impressive performance metrics compared to state-of-the-art models like GPT-4 and Claude 3
- Autonomy in decision-making for using external tools in the GLM-4 All Tools model
- Practical applications include web browsing, Python interpretation, accessing online information, and solving math problems
- Open-sourcing of various models within ChatGLM family with over 10 million downloads on platforms like Hugging Face
- Commitment to promoting accessibility and safety of Large Language Models through open releasing model weights and techniques
- Continuous refinement based on lessons learned from previous generations
- Democratizing cutting-edge LLM technologies through open sourcing efforts to push boundaries towards teaching machines to think more like humans
- Acknowledgments to contributors at Zhipu AI, Tsinghua University, collaborators, partners, and organizations supporting open-sourcing efforts
- Represents a significant step forward in understanding and executing complex tasks autonomously in natural language processing

## Summary

In this report, we introduce the ChatGLM family of large language models. Specifically focusing on the GLM-4 series which includes GLM-4, GLM-4-Air, and GLM-4-9B. These models have been trained on a vast amount of data in multiple languages with a strong emphasis on Chinese and English. Through a multi-stage post-training process involving supervised fine-tuning and human feedback, the GLM-4 models have achieved impressive performance metrics compared to state-of-the-art models like GPT-4 and Claude 3. One key advancement in the GLM-4 All Tools model is its ability to autonomously decide when and which external tools to use for complex tasks such as web browsing or Python interpretation. This capability has enabled the model to excel in practical applications like accessing online information and solving math problems. Furthermore, the open-sourcing of various models within the ChatGLM family has garnered significant interest with over 10 million downloads on platforms like Hugging Face. The commitment to promoting accessibility and safety of Large Language Models (LLMs) through open releasing model weights and techniques reflects a dedication to advancing LLM technologies while ensuring transparency. Looking ahead, the team is continuously refining their models based on lessons learned from previous generations. By democratizing cutting-edge LLM technologies through open sourcing efforts, they aim to push the boundaries of model capabilities towards teaching machines to think more like humans. Acknowledgments are extended to all those who have contributed to the development of ChatGLM models at Zhipu AI and Tsinghua University as well as collaborators and partners who have supported this journey. Special thanks are given to individuals from various organizations who have assisted in open-sourcing efforts. Overall, the ChatGLM family of large language models represents a significant step forward in understanding and executing complex tasks autonomously. With ongoing advancements and a commitment to openness and collaboration, these models are poised to continue making strides in the field of natural language processing.