GLM-4.5V

GLM-4.5V is a vision-language foundation model built for multimodal agent applications. Based on a Mixture-of-Experts architecture with 106B total parameters and 12B activated parameters, it is engineered to excel at tasks that require understanding across visual and textual boundaries. The model demonstrates state-of-the-art performance in video understanding, image Q&A, OCR, document parsing, and particularly in front-end web coding, grounding, and spatial reasoning tasks. A distinctive feature is its dual-mode inference approach, allowing developers to toggle between deep "thinking mode" for complex reasoning chains and a fast "non-thinking mode" for straightforward responses, making it adaptable across different agentic use cases.

GLM-4.5V continues the technical lineage of GLM-4.1V-Thinking, incorporating scalable reinforcement learning methods to expand reasoning capabilities beyond basic multimodal perception. The model achieves state-of-the-art results among models of its scale across 42 public vision-language benchmarks, reflecting rigorous development investment. Its open weights availability enables developers to fine-tune and deploy it for specialized agent applications ranging from document intelligence pipelines to GUI automation. The togglable reasoning behavior makes it practical for workflows requiring both deep analytical chain-of-thought and rapid response generation depending on task complexity.