OpenBMB's first OCR-specific models, split from the MiniCPM-V line and released under Apache 2.0 (weights on HuggingFace from October 21, the GitHub launch on October 27, 2026). MiniCPM-OCR 1.0 is a 1.3B document parser built on MiniCPM-V 4.6 that turns Chinese and English documents and PDFs into text, tables, formulas, charts and seals, including scans, photos, curved and recaptured pages. OpenBMB reports state-of-the-art open results of 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2, with leading scores on OmniDocBench v1.6, Wild-OmniDocBench and olmOCR-Bench, at about 1.2 pages per second, which it puts at 1.49× PaddleOCR-VL-1.6's throughput and 3.62× MinerU2.5 Pro's. A DSpark draft head adds a further 55% end-to-end speedup.

The more unusual release is MiniCPM-OCR Jev, a 2B OCR correctness decision model built on Qwen3.5-2B. Given a region image and a candidate transcription from any OCR system, it returns a 0–1 score for whether the transcription is right, so pipelines can auto-accept reliable regions and route the rest to review. On OpenBMB's new 2,267-record JevOCRBench it reaches an AUROC of 93.6, against 90.1 for a 35B-A3B judge and 93.8 for Qwen3.5-397B-A17B, and it holds 91.9–93.4 when scoring output from MiniCPM-OCR, GLM-OCR, PaddleOCR-VL-1.6 and Xiaomi-OCR-0. All figures are OpenBMB's own.

Model Details

Architecture DENSE
Parameters 1.3B
License Apache 2.0
Base model minicpm-v-4.6

Benchmark Scores

Benchmark Score Mode
Real5-OmniDocBench 93.68 —
OCRBench v2 54.7 —

Variants

Name Parameters Notes
MiniCPM-OCR-1.0 1.3B Document parser from MiniCPM-V 4.6; Real5-OmniDocBench 93.68, OCRBench v2 54.7; DSpark and GGUF builds
MiniCPM-OCR-Jev 2B OCR correctness decision model from Qwen3.5-2B; JevOCRBench AUROC 93.6
ocrvisionopen-weighton-device

Related