GLM-OCR is a multimodal optical character recognition model designed for complex document understanding. Built on the GLM-V encoder-decoder framework, it introduces Multi-Token Prediction (MTP) loss and full-task reinforcement learning to enhance training efficiency and accuracy. The model uses a CogViT visual encoder pre-trained on extensive image-text datasets, along with a lightweight cross-modal connector and a GLM-0.5B language decoder. Its two-stage pipeline, featuring layout analysis and parallel recognition, allows it to perform well across various document layouts. In terms of performance, GLM-OCR achieved a score of 94.62 on OmniDocBench V1.5, ranking first overall among document understanding benchmarks. It excels in tasks like formula recognition, table recognition, and information extraction, making it particularly useful for real-world applications that involve complex layouts. With just 0.9 billion parameters, it supports various deployment methods, reducing latency and costs, which is crucial for high-concurrency services. The official SDK simplifies usage by integrating layout analysis and structured output generation, minimizing the engineering effort needed for document intelligence systems. Users can get started quickly with commands for installing dependencies, running models, and executing tasks through different platforms like vLLM and SGLang. GLM-OCR currently supports document parsing and structured information extraction, requiring specific JSON formats for the latter to ensure compatibility in downstream processing.