OpenBMB Open-Sources MiniCPM-OCR Suite, 1.3B Model Hits OCR SOTA
On October 28, OpenBMB released and open-sourced its first OCR models, the MiniCPM-OCR series, now available on Hugging Face under the Apache 2.0 license. The series includes two models: the parsing model MiniCPM-OCR 1.0 (1.3B parameters) and the quality-judge model MiniCPM-OCR Jev (2B parameters).
Confirmed
- MiniCPM-OCR 1.0 is a Chinese-English document parsing model built for real-world scenarios such as curved pages, phone photos, skewed scans, complex layouts, tables, and formulas, scoring 93.68 on Real5-OmniDocBench (m1, m3, m4)
- The team says it achieves SOTA-level parsing speed and accuracy at just 1B scale, outperforming solutions from Xiaomi and PaddleOCR in benchmarks (as reported in m3)
- Combined with DSpark decoding acceleration, end-to-end processing efficiency improves by 55% (m4, m5)
- The official efficiency-performance scatter plot shows the model leading across four benchmarks (m4)
- MiniCPM-OCR Jev (2B) takes an image crop and its OCR output and produces a 0-1 quality score covering text, formulas, and tables; it can also score results from other OCR parsers (m2 and m4 mention an AUROC of 0.936)
Why it matters
- The small 1.3B parameter count plus DSpark acceleration allows high-accuracy document parsing to run on-device or in low-cost environments, lowering the barrier to OCR deployment
- The built-in, reusable quality-judge model offers a path to automated OCR quality inspection and isn't tied to the team's own parser, giving it strong ecosystem compatibility
- The Apache 2.0 license permits commercial use, making this a directly adoptable open-source component for downstream applications like document processing and RAG
2026-10-28 ~ 2026-10-28 · 7 related posts
Primary sources
- OpenBMB ships 1.3B OCR model scoring 93.68 plus a 2B AI judge for OCR errors — Equivalent-Grass-527 ·
- OpenBMB ships 1.3B MiniCPM-OCR 1.0 plus a 2B verifier model, Apache 2.0 — BerryBlush_Rogers691 ·
- OpenBMB open-sources MiniCPM OCR family: 1.3B model tops open-source OCR benchmarks, 2B evaluator hits 0.936 AUROC — AdinaYakup ·
- [source] OpenBMB ships 1.3B OCR model scoring 93.68 plus a 2B AI judge for OCR errors — Equivalent-Grass-527 · 2026-10-28
- OpenBMB releases open-source MiniCPM-OCR 1.0 with online demo — Dr_Singularity · 2026-10-28
5 near-duplicate retellings: BerryBlush_Rogers691 · AdinaYakup · AdinaYakup · vanstriendaniel · AIwithGhotai