About
The GLM-4 family, developed by Zhipu AI and Tsinghua University, represents an evolving series of large language models renowned for their multilingual capabilities and state-of-the-art performance. Building upon previous ChatGLM generations, these models are pre-trained on an extensive dataset of ten trillion tokens across Chinese, English, and 24 other languages. They undergo rigorous multi-stage post-training, involving supervised fine-tuning and reinforcement learning from human feedback, which enables them to rival or surpass GPT-4 on various benchmarks. The series includes versions like GLM-4, GLM-4-Air, and GLM-4-9B, each tailored for different tasks and resource constraints. A notable feature is the GLM-4 All Tools model that can autonomously use web browsers and Python interpreters for complex task completion. Open-source variants, such as GLM-4-9B and its chat-optimized version, along with multimodal models like GLM-4V-9B, which integrates image processing, highlight the family's versatility. Recent advancements include the GLM-4-Voice model, an end-to-end speech model supporting Chinese and English, further extending the boundaries of open-source LLMs 1 3 5 6 7 8.