Mianbi Intelligence open-sources the MiniCPM-OCR series, covering recognition and result verification

Article

ModelBest Open-Sources OCR Models: 1.3B One for recognition, 2B the other specialized in error detection

Beating AI News: ModelBest has released and open-sourced the MiniCPM-OCR series. 1.3B The MiniCPM-OCR 1.0 can recognize handwritten text, complex tables, formulas, and skewed or curved documents. Across OCR benchmark tests, it ranks first in.RTX 5090 achieves a processing speed of 1.217 pages per second, about PaddleOCR-VL-1.6 faster than 49%.

The other 2B model, MiniCPM-OCR Jev, is specifically responsible for error detection, checking OCR results against the original image and assigning separate scores for text, formulas, and tables. High-scoring content can pass directly, while questionable content is then submitted for manual review. It can also check GLM-OCR、Paddle OCR the output of other models and does not require replacing the existing OCR system.

This design targets the most labor-intensive stage of the OCR workflow: machines can read documents quickly, but staff still have to check item by item for missing characters, incorrect lines, or number-recognition errors. ModelBest's internal experiments on 900 manually annotated image patches show that when the accuracy of automatically approved portions approaches 100%, up to 85%of areas in legal documents require no manual verification; for financial invoices, the figure is 56.2%, and for complex documents, 35.1%.

Original link https://m.theblockbeats.info/flash/374326

Key points

Mianbi Intelligence releases and open-sources the MiniCPM-OCR series.
1. 3B’s MiniCPM-OCR 1.0 can recognize handwritten text, complex tables, and formulas, and can also process documents that are photographed at an angle or are curved.
It was reported that MiniCPM-OCR 1.0 ranked first in four of the six OCR benchmarks.
On RTX 5090, MiniCPM-OCR 1.0 has a processing speed of 1.217 pages per second, compared with PaddleOCR-VL-1.6 faster by approximately 49%。
2B model MiniCPM-OCR Jev can compare the results against the original image OCR and score text, formulas, and tables.
MiniCPM-OCR Jev can inspect the outputs of GLM-OCR、Paddle OCR and other models.
ModelBest's internal experiments were based on 900 manually annotated image blocks; it was reported that when the accuracy of the automatic pass-through portion approached 100%, up to 85% of regions in legal documents required no manual verification, compared with 56.2% for financial invoices and 35.1%。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This solution divides OCR the work into two steps: the first model reads the document, and the second model then checks the recognition results against the original image. High-confidence content can pass directly, while questionable content is referred for manual review.

Why it matters to readers

If the relevant performance and internal experimental results can be independently reproduced,OCR manual item-by-item checks in the process may be reduced; the review model can also check the outputs of other OCR models.

OCR It converts text, tables, and formulas in images or scanned documents into processable information. The second model here functions as an automated quality inspector.

Risks and unknowns

It is not stated whether the benchmark tests and performance data in the text can be independently reproduced.
900 The experiment on manually annotated image tiles was an internal closed-door experiment; whether the results are representative of real-world deployment scenarios remains unclear.
The evidence does not provide the model code, weights, license, or information on actual deployment costs.

Mianbi Intelligence open-sources the MiniCPM-OCR series, covering recognition and result verification

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center