OpenBMB/MiniCPM-OCR

MiniCPM-OCR is a family of released, open-weight models for turning document images and PDFs into structured text. It targets teams that need OCR they can run or fine-tune themselves, not a hosted document-processing product. The repository ships weights and guidance rather than an end-user app.

Two models for two jobs

MiniCPM-OCR 1.0 (1.3B parameters) parses Chinese and English documents and PDFs into Markdown or JSON, covering text, tables, formulas, charts, seals and key-information extraction. The authors report a 93.68 score on Real5-OmniDocBench and describe handling of scans, photographs, curved, skewed and recaptured pages README. Higher throughput is claimed with optional DSpark decoding heads (BF16 and GGUF variants), which the authors say improves end-to-end efficiency by 55% DSpark weights. Those efficiency figures are author-reported benchmark conditions; see the throughput documentation for settings throughput notes.

MiniCPM-OCR Jev (2B) is a separate correctness-decision model: given an image and its OCR result, it returns a quality score from 0 to 1, so regions needing review can be separated from reliable ones. The authors report AUROC (×100) of 93.6 on their JevOCRBench and 91.9–93.4 when scoring outputs from four parsing models, including MiniCPM-OCR 1.0 and third-party OCRs Jev guide. This is a model-evaluation claim, not a guarantee that every bad region is caught.

Ways to use it

• Hosted demo and free API for recognition plus Jev scoring, without local hardware API example; the Jev API example is listed as coming soon.
• Local deployment of either model, with tutorials for parsing and for scoring deployment tutorial.
• Serving through vLLM, SGLang, llama.cpp or Ollama, with BF16, GGUF, AWQ and GPTQ weight variants published on Hugging Face and ModelScope model weights.
• Fine-tuning via LLaMA-Factory or Swift for the parser, and Swift for Jev fine-tuning docs.
• Coding-agent skill: an agent can follow the provided SKILL file to set up an environment, download models and parse a first document SKILL.

The pip SDK offers another integration path pip package.

Prerequisites and limits

Local use requires downloading model weights and adequate GPU memory; the FAQ addresses installation, downloads, GPU memory, input formats and unexpected outputs FAQ, but the README does not state a minimum memory figure. The parser is described for Chinese and English, so other languages are outside the stated scope. Benchmark results are author-reported on named benchmarks and should be read with their reproduction settings reproduction guide. Free-API availability and quota terms are not specified in the supplied material.

Code, documentation and the weights of both models are released under Apache-2.0 license, so commercial and modification use is permitted subject to that licence. There are no hands-on test results here; suitability for a specific document mix should be checked against the guides and on your own samples.

HotRepo's take

HotRepo recorded this project at #22 on its NEW-REPOSITORY board with 197 total stars on 2026-10-28; that is a single dated snapshot, not a growth trend. The evidence supports a model-release project worth evaluating on your own documents, not a judgement about reliability or adoption.