GLM Models by Tsinghua Knowledge Engineering Group (THUDM)

About

The GLM family of large language models (LLMs) is a collaborative creation by the GLM team at Zhipu AI and Tsinghua University 1. These models range from 110 million to a massive 130 billion parameters and are especially notable for their bilingual proficiency, supporting both English and Chinese languages 7. The GLM architecture is distinct for its use of autoregressive blank infilling as the core pre-training strategy 10. Among these, the GLM-130B model stands out by matching or surpassing the performance of GPT-3 in various benchmarks, even outdoing ERNIE TITAN 3.0 in Chinese language tasks 7. Successive models like GLM-4 have been trained on ten trillion tokens and employ refined techniques like supervised fine-tuning and reinforcement learning from human feedback, enhancing alignment and instruction adherence 1. The GLM-4 series, including the GLM-4 All Tools, showcase the ability to understand user intent and use tools autonomously, such as web browsers and Python interpreters, for executing intricate tasks 1.

Specifications(2 models)

Related Model Families

ChatGLM-4

GLM-Z1

CogVLM2

ChatGLM3

Models(2)