GLM-4.6V is the latest release in open-source multimodal large language models, featuring two versions: the 106 billion parameter foundation model for high-performance environments and the 9 billion parameter Flash version for local deployment. It boasts a significant increase in context length, handling up to 128k tokens, and achieves state-of-the-art performance in visual understanding and reasoning. The model introduces native function calling, allowing it to process visual inputs like images and documents directly, without prior conversion to text. This integration reduces information loss and simplifies the workflow for complex tasks. The model excels in several applications. It can generate structured content from diverse inputs, like reports and presentations, while autonomously invoking tools to crop visuals and assess their relevance. For visual web searches, GLM-4.6V identifies user intent, retrieves pertinent information, and synthesizes it into coherent answers. In frontend development, it streamlines the transition from design to code, generating high-fidelity HTML/CSS/JS from screenshots and allowing for interactive modifications through natural language commands. GLM-4.6V also supports extensive context processing, making it effective for analyzing long documents or videos. For example, it can simultaneously extract key metrics from financial reports of multiple companies and summarize lengthy video content while preserving detailed reasoning on specific events. Rigorous evaluations across 20 multimodal benchmarks demonstrate its leading capabilities in multimodal understanding and logical reasoning. The underlying techniques include training on a billion-scale multimodal dataset to enhance visual perception and cross-modal question answering. The model employs advanced methodologies for handling multimodal outputs and incorporates reinforcement learning to refine its task planning and tool invocation skills. This development represents a significant leap forward in creating models that can seamlessly integrate visual and textual information for practical applications.