MiniCPM-OCR-1.0: a 1B Apache-licensed OCR model that runs local
OpenBMB's doc model hits 1.2 pages/sec on one RTX 5090, beating PaddleOCR-VL and MinerU — a fully local, no-API-bill parser you can wire in this weekend.
Copy markdown
1B params, Apache 2.0, and it runs in Ollama
OpenBMB released MiniCPM-OCR-1.0, a ~1B OCR-focused vision-language model under Apache 2.0 that reads text, tables, formulas, charts and seals and handles key-information extraction. It reports 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2, and pushes about 1.2 pages/sec on a single RTX 5090 — roughly 1.5x PaddleOCR-VL-1.6 and 3.6x MinerU2.5 Pro — with a `minicpmvocr` pip SDK that parses whole PDFs end to end by pairing layout detection with the model. It runs in vLLM, SGLang, Ollama, llama.cpp and Transformers, with AWQ/GPTQ/GGUF quants, so a no-API-bill document pipeline is now a genuine weekend build. It's tuned mainly for Chinese and English.