Is MiniCPM-OCR Jev state of the art?Pending benchmarks.Announced Oct 27, 2026; no independent results yet.
OpenBMB's OCR checking model. Given a text, formula or table region from a document and a candidate OCR result from any parser, it returns a score from 0 to 1 for how likely the result is correct, to auto-accept reliable regions, send the rest to human review, or pick among several OCR models. Open weights under Apache 2.0.
The details so far
- Lab
- OpenBMB
- Announced
- Oct 27, 2026
- API
- Not yet · Cloud API listed as coming soon in the lab's GitHub repo
- Open weights
- Published · Hugging Face openbmb/MiniCPM-OCR-Jev, plus GGUF, AWQ and GPTQ builds
- Parameters
- 2B
- Base
- Trained from Qwen3.5-2B; vision encoder and language backbone kept, LM head replaced with a binary classification head
- Input / output
- Region image + candidate OCR text (and its rendered image) in; a 0-1 correctness score out
- Languages
- Chinese and English
- License
- Apache 2.0
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 93.6 AUROC (x100) on JevOCRBench, above JEV-35B-A3B (90.1) and close to Qwen3.5-397B-A17B (93.8); 91.9-93.4 on parses from MiniCPM-OCR 1.0, GLM-OCR, PaddleOCR-VL-1.6 and Xiaomi-OCR-0.
Sources
LABOpenBMB announcement on Xx.com/OpenBMB/status/2115124474882469955 · Oct 27, 2026REPOOpenBMB/MiniCPM-OCR on GitHubgithub.com/OpenBMB/MiniCPM-OCR · Oct 27, 2026REPOopenbmb/MiniCPM-OCR-Jev on Hugging Facehuggingface.co/openbmb/MiniCPM-OCR-JevTracking continues until an independent board lists it.