GLM-OCR: Zhipu AI's 0.9B Model Tops OmniDocBench V1.5 — Beats Models 260x Its Size

GitHub·medium signal

Zhipu AI and Tsinghua University open-sourced GLM-OCR, a 0.9B parameter multimodal OCR model scoring 94.62 on OmniDocBench V1.5, beating Qwen3-VL-235B (260x larger). Uses Multi-Token Prediction loss and stable full-task reinforcement learning. Processes 1.86 pages/sec for PDFs. Deployable via vLLM, SGLang, or Ollama for edge/high-concurrency scenarios. MIT licensed. At 5.2K stars and trending at +238/day.