MiniCPM-OCR 1.0: OpenBMB releases a small OCR model under Apache 2.0, paired with a second model that scores result reliability

OpenBMB, the open-source collective behind the MiniCPM model family, published the MiniCPM-OCR suite on Tuesday, October 27, announced on X at 5:52 PM Paris time. It brings together two models with downloadable weights under the Apache 2.0 license: MiniCPM-OCR 1.0, which reads and structures documents, and MiniCPM-OCR Jev, which checks the first model’s work.

The idea is simple to state: an OCR (optical character recognition) engine that “finishes” a page doesn’t say whether it made a mistake. OpenBMB therefore offers a fast reader and a controller capable of flagging doubtful passages.

A 1.3-billion-parameter document reader

According to its Hugging Face card, MiniCPM-OCR 1.0 has 1.3 billion parameters. It is the first OCR-dedicated model in the MiniCPM-V family: it is trained from MiniCPM-V 4.6, which combines a SigLIP visual encoder and a Qwen3.5 language model. It recognizes text, tables, formulas, charts, and stamps, and can extract specific fields (a store name and a receipt amount, for example) as JSON.

The model does not read a page in one block. The development kit first cuts up the document with a layout model (PP-DocLayoutV3), sends each region to the model, then reassembles everything into Markdown in reading order, with tables in HTML and formulas in LaTeX. The priority languages are Chinese and English; the documentation nonetheless shows an example of fine-tuning on handwritten French lines.

A second model that scores each passage

MiniCPM-OCR Jev (2 billion parameters, derived from Qwen3.5-2B) does not read documents: it receives the image of a region and the text an OCR engine extracted from it, and returns a score from 0 to 1, all the higher as the result is likely to be correct. OpenBMB replaced the language model’s output head with a binary classification head.

Three uses are described: automatically accepting regions above a threshold, sending the others to human review, or having the same region read by several OCR engines and keeping the best result. Jev does not depend on MiniCPM-OCR: the developer tested it on the outputs of GLM-OCR, PaddleOCR-VL-1.6, and Xiaomi-OCR-0. The threshold remains to be set by the user, on their own documents, as the card specifies.

Numbers to verify

All published results come from OpenBMB. MiniCPM-OCR 1.0 would score 93.68 on Real5-OmniDocBench, a test of real documents (scanned, curved, photographed, or skewed pages), and 54.7 on OCRBench v2. On speed, the developer announces 1.2 pages per second on an NVIDIA GeForce RTX 5090 card, i.e. 1.49 times PaddleOCR-VL-1.6 and 3.62 times MinerU2.5 Pro, and up to 128 pages per minute by parallelizing processing.

For Jev, the area under the ROC curve (AUROC, ×100), which measures the ability to separate good from bad results, would reach 93.6 on JevOCRBench, above JEV-35B-A3B (90.1) and close to Qwen3.5-397B-A17B (93.8), a much larger model. But this benchmark is the developer’s, and the test on four OCR engines covers only 200 samples.

Ready to deploy

Both models work out of the box with vLLM, SGLang, llama.cpp, and Ollama, exist in quantized versions (GGUF, AWQ, GPTQ), and can be fine-tuned. OpenBMB cites day-one support from Arm, Houmo AI, Intel, and NVIDIA, and other chip makers via FlagOS. An online demo and an API are also accessible on the ModelBest platform, to which the announcement links.