ChatGLM2 6B
ChatGLM2-6B is an open-source bilingual conversational large language model developed by Tsinghua University, serving as the second-generation successor to ChatGLM-6B. It features enhanced performance with significant improvements on benchmarks like MMLU, CEval, and GSM8K achieved through hybrid objective training on 1.4T tokens. The model supports a 32K context length via FlashAttention technology, enabling efficient handling of long documents and multi-turn dialogues. It offers optimized inference speed and reduced memory usage using Multi-Query Attention, allowing INT4 quantization to run on 6GB GPU memory. ChatGLM2-6B is designed for both research and free commercial use after simple registration. It serves as a foundation for various derivative projects including code generation variants like CodeGeeX2 and long-context specialized versions. The model is compatible with acceleration frameworks such as fastllm and chatglm.cpp for deployment on diverse hardware including mobile devices and TPU chips.