Skip to main content
Back

OpenBMB's MiniCPM-OCR Pairs a 1.3B Parser With a 2B Accuracy Verifier

OpenBMB released a two-model OCR system: a 1.3B parser hitting state-of-the-art accuracy and a 2B verifier that scores whether its output is correct.

  • MiniCPM-OCR 1.0 is a 1.3B document parser scoring 93.68 on Real5-OmniDocBench.
  • MiniCPM-OCR Jev is a 2B verifier that scores OCR candidate correctness from 0 to 1.
  • Jev hits 93.6 AUROC, matching a 397B model while being nearly 200x smaller.
  • Jev generalizes across parsers including GLM-OCR, PaddleOCR-VL, and Xiaomi-OCR with 91.9 to 93.4 AUROC.
  • Both ship Apache 2.0 with vLLM, SGLang, llama.cpp, Ollama, and quantized GGUF/AWQ/GPTQ variants.
  • Use cases: auto-accept with human-review routing, and best-of-N selection across multiple OCR models.

OpenBMB and ModelBest have released two Apache 2.0 models for production OCR pipelines. The parser checkpoint extracts text and document structure, while the Jev checkpoint scores whether each extracted region is correct. The pairing gives developers a way to automate acceptance, route uncertain results to reviewers, or choose among outputs from several parsers.

A 1.3B parser takes the first pass

MiniCPM-OCR 1.0 builds on MiniCPM-V 4.6 and combines a SigLIP vision encoder with a Qwen3.5 language backbone. The vision encoder converts page images into representations that the language model uses to recover text, tables, formulas, charts, seals, and structured fields. Chinese and English are its primary languages.

Its reported benchmark and throughput results include:

  • Real5-OmniDocBench: An aggregate score of 93.68 on documents that include scans, photographs, curved pages, skew, and recaptured images.
  • Additional benchmarks: Leading results reported on OmniDocBench v1.6, Wild-OmniDocBench, and olmOCR-Bench.
  • DSpark acceleration: A reported 55% improvement in end-to-end efficiency.
  • Relative throughput: About 1.49 times PaddleOCR-VL-1.6 and 3.62 times MinerU2.5 Pro under the project’s test settings.
  • RTX 5090 throughput: Approximately 1.2 pages per second.

The figures come from the project’s release materials, and throughput will vary with page complexity, backend, concurrency, image resolution, and hardware. The benchmark’s concurrency setting of one applies at the image level; regions within a page are still recognized in parallel.

Jev turns OCR checking into classification