Chinese AI startup Zhipu AI, in collaboration with ByteDance and Tsinghua University, has released the technical report for GLM-4.5, an open-source large language model (LLM) designed for agentic, reasoning, and coding tasks. The model employs a unique multi-stage training paradigm featuring expert model iteration with self-distillation to unify capabilities across these areas. GLM-4.5 uses a Mixture-of-Experts (MoE) architecture and was trained on 23 trillion tokens with reinforcement learning enhancements. It ranks third overall and second on agentic tasks across 12 benchmark tests. Building on this, Zhipu AI introduced GLM-4.5V, a vision-language model that extends GLM-4.5’s capabilities to visual reasoning, achieving state-of-the-art performance on 41 to 42 benchmarks covering image, video, and document understanding. GLM-4.5V is based on the GLM-4.5-Air base model and features a 106 billion-parameter MoE architecture. Both models are available on platforms such as Hugging Face and Anycoder, supporting tasks that range from general language understanding to advanced visual analysis. The release highlights ongoing advancements in open-source AI models with strong performance in hybrid reasoning and agentic functionalities.