MonkeyOCR

A structure-recognition-relation triplet paradigm sets MonkeyOCR apart from chained document pipelines: one unified model detects layout structure, recognizes content, and predicts relationships between elements, emitting markdown plus annotated PDFs and per-block JSON with coordinates. It ships in 1.2B and 3B parameter variants; the pro 3B model outperforms GPT-4o, Gemini 2.0 Flash, Qwen2.5-VL-72B, and InternVL3-78B on OmniDocBench across both English and Chinese documents, while the 1.2B version runs about 36 percent faster for roughly 1.6 percent lower accuracy. Hardware needs are modest for the class: the project reports successful runs on an 8 GB RTX 4060, the RTX 3090 and 4090, the A6000, A100, and H800, with throughput between roughly 0.5 and 2.4 pages per second depending on hardware and model size. Installation is pip with CUDA support, with weights pulled from Hugging Face or ModelScope. The code is Apache-2.0, but the weights are limited to academic research and non-commercial evaluation, with commercial use requiring a separate agreement. The repository has about 6,600 stars.