Chinese AI startup Zhipu AI, in collaboration with ByteDance, has released the technical report for GLM-4.5, a new open-source large language model (LLM) family designed to excel in agentic, reasoning, and coding (ARC) tasks. The model employs a unique multi-stage training paradigm, including expert model iteration with self-distillation and reinforcement learning, trained on 23 trillion tokens. GLM-4.5 features a 355 billion-parameter Mixture-of-Experts (MoE) architecture that enables efficient scaling and unification of specialized expert models into a single versatile foundation model. It ranks third overall across 12 benchmarks and second on agentic tasks, demonstrating strong generalization capabilities such as web search and software engineering tasks. Additionally, Zhipu AI introduced GLM-4.5V, a vision-language variant built on the GLM-4.5-Air base model. GLM-4.5V uses a 106 billion-parameter MoE architecture and inherits advanced reasoning techniques from the earlier GLM-4.1V-Thinking model. It achieves state-of-the-art performance on 41 to 42 benchmarks covering image, video, and document understanding and is available on platforms including Hugging Face and AnyCoder. The release marks a notable advancement in open-source models capable of handling complex multi-modal and agentic applications.