Zai/GLM-OCR (huggingface.co)

Zai has introduced GLM-OCR, a groundbreaking multimodal Optical Character Recognition (OCR) model designed for complex document understanding. Built on the GLM-V encoder–decoder architecture, this model incorporates innovative features like Multi-Token Prediction (MTP) loss and full-task reinforcement learning to enhance training efficiency and recognition accuracy. With a combination of the CogViT visual encoder and a lightweight cross-modal connector, GLM-OCR excels in handling diverse document layouts, achieving state-of-the-art performance with an impressive score of 94.62 on the OmniDocBench V1.5 benchmark. This model is particularly significant for the AI/ML community because it addresses real-world challenges in document processing, such as complex tables and document formats that are traditionally difficult for OCR technologies. Optimized for deployment in high-concurrency environments, GLM-OCR boasts only 0.9 billion parameters, minimizing inference latency and computational costs. Furthermore, Zai has made the model fully open-source, complemented by a comprehensive SDK for easy integration into existing production workflows, making robust OCR capabilities accessible for businesses and developers alike.