zai-org/GLM-OCR at 6.5K Stars: 0.9B-Param OCR Model Ranks #1 on OmniDocBench V1.5 With 94.62 Score

GitHub·medium signal

GLM-OCR combines CogViT visual encoder with GLM-0.5B language decoder in just 0.9B parameters, achieving 94.62 on OmniDocBench V1.5 to rank #1 overall. Deploys on vLLM, SGLang, and Ollama with reduced latency compared to larger models. Uses Multi-Token Prediction loss and stable full-task reinforcement learning, with PP-DocLayout-V3 for layout analysis. Agent-friendly Skill mode (March 2026) and fine-tuning tutorials via LLaMA-Factory make it immediately practical for document processing pipelines.