This repository contains the GLM-4 model series, open-source multilingual multimodal chat language models. The models include GLM-4-32B-0414 with 32 billion parameters trained on 15T tokens of high-quality data including synthetic reasoning data. The series supports local deployment and uses reinforcement learning, rejection sampling, and human preference alignment for training and alignment. Performance benchmarks show results comparable to GPT and DeepSeek V3/R1 models.