MiniCPM-OCR Is Two Small Open Models: One Reads Documents, the Other Checks Its Work

ModelBest and OpenBMB released MiniCPM-OCR yesterday, 31 October 2026: two models with two jobs. MiniCPM-OCR 1.0, at 1.3 billion parameters, reads documents. MiniCPM-OCR Jev, at 2 billion parameters, does not read documents at all. It checks the reading.

Both models are listed on GitHub, Hugging Face, and ModelScope, with four runtimes named in the release: vLLM, SGLang, llama.cpp, and Ollama. The release contains no benchmark scores and states no licence type. The scores and licence below come from the models' Hugging Face cards, checked today — the vendors' own figures, not independent results.

MiniCPM-OCR 1.0 turns messy pages into structured text

Picture the input: a scanned contract with a curved page, a photographed receipt, tilted, with a shadow across one corner, a screenshot of a table, a PDF printed, signed, and scanned back in. According to the release, MiniCPM-OCR 1.0 parses scanned, photographed, curved, tilted, and screen-captured documents without what the vendors call "cumbersome image pre-correction" — the straightening and cleaning older pipelines needed first.

Parsing means more than reading characters. The model must recover structure: where a paragraph ends, where a table begins, what belongs in which cell, and how a formula should be written so it can be used again. The release says MiniCPM-OCR 1.0 handles dense bilingual Chinese-English content, complex mathematical formulas, and multi-level or borderless tables, with full-page reconstruction. The card adds charts, seals, and structured extraction to the outputs.

It also supports key information extraction, or KIE: pulling specified fields on instruction, with invoices, IDs, and financial statements as examples. A scanned invoice returned as a wall of text still needs a person to find the amount, date, and vendor. One returned as structured fields can feed an accounting system directly. Whether it does that reliably on a given organisation's documents is a question no announcement can answer.

Illustration created for INORYVA — a document being scanned and read, not a screenshot of MiniCPM-OCR output.

MiniCPM-OCR Jev checks the work instead of doing it

Jev is the part most likely to be misdescribed, so the distinction needs stating plainly. MiniCPM-OCR Jev is not an OCR model. It does not read documents. It is what the release calls the first OCR result decision model, acting as an "AI quality inspector." Given a source image and a candidate OCR result, Jev compares the two and assigns a confidence score — on a 0-to-1 scale on the Hugging Face card, where higher means more likely correct. It scores text, LaTeX formulas, and tables, and according to the release detects missing characters, errors, and formatting misalignments.

That answers a specific problem. Organisations digitising millions of pages — the release names financial contract archives, academic libraries, and government records — have needed page-by-page manual review, because one unchecked error can invalidate a record. Jev is meant to replace blanket review with risk-based review: reliable results go straight through, doubtful ones are flagged for a human, and where several OCR systems read the same page, Jev can select the best result. It works across models, without replacing the underlying OCR engine.

Jev is trained from Qwen3.5-2B, keeping the vision encoder and language backbone but replacing the language-model output head with a binary classification head that learns whether a candidate is correct: a specialist trained for one yes-or-no question per region of a page. A threshold decides the workflow — results above it proceed automatically, the rest go to human review — and the card warns the threshold should be chosen from samples of the actual application.

Illustration created for INORYVA — structured text and tables extracted from a document, not a screenshot of MiniCPM-OCR output.

The vendors' claims, fenced — and what the model cards add

The release's quality language is vendor wording, reported that way. ModelBest and OpenBMB describe MiniCPM-OCR 1.0 as bringing "top-tier parsing accuracy" to a compact 1.3-billion-parameter architecture, performing at a "leading level across mainstream industry evaluations." Neither phrase is an independent finding, and the release text contains no named benchmark, score, or comparison table to check either claim against.

The Hugging Face cards, checked on 1 November 2026, do publish figures — read them as the vendors' case, not verified fact. MiniCPM-OCR 1.0 is reported at 93.68 on Real5-OmniDocBench, a real-world document-parsing benchmark, with leading results also claimed on OmniDocBench v1.6, Wild-OmniDocBench, and olmOCR-Bench, plus 54.7 on OCRBench v2. Throughput is reported at 1.2 pages per second on an NVIDIA GeForce RTX 5090, across 1,651 OmniDocBench v1.6 images. Scores from different benchmarks are not directly comparable — the card says so itself — and none had been independently reproduced in any source checked. DSpark acceleration is described in the release as significantly increasing throughput while lowering latency; the card quantifies it as a 55% efficiency improvement, with throughput about 1.49 times PaddleOCR-VL-1.6 and 3.62 times MinerU2.5 Pro, in the vendors' own setup. That is not a general speed guarantee.

Jev's figures follow the same pattern. The Jev card reports a decision AUROC of 93.6 (scaled by 100) on JevOCRBench — the vendors' own benchmark — above a 35B model at 90.1 and close to a 397B model at 93.8, and 91.9 to 93.4 on parses from four different OCR systems in a 200-sample experiment. A small model outscoring far larger ones on a narrow decision is plausible, since judging is narrower than reading, but the benchmark and comparisons are the vendors', and the card advises validating separately on target data. The honest reading: Jev looks promising on the vendors' own tests, and teams should validate it on their own documents before trusting auto-acceptance. No comparison with other open-weight releases is made on invented benchmarks. This small specialist plays a different role from the flagship generalists in INORYVA's flagship AI model comparison: one job, on documents, at a size that can be deployed widely.

How to get it and run it

The release lists GitHub, Hugging Face, and ModelScope, and the cards confirm the weight repositories, including GGUF builds for llama.cpp, AWQ and GPTQ quantised variants, and DSpark draft weights that must be matched with a main checkpoint from the same release. Runtime support is vLLM, SGLang, llama.cpp, and Ollama, alongside software development kits, APIs, Agent Skill plugins, and fine-tuning code. The cards add an online demo and a recommended local route: the vendors' software kit with DSpark enabled by default, reading a PDF or image, detecting layout, recognising each region, and assembling Markdown in reading order.

The licence, unstated in the release, is confirmed on both cards: Apache License 2.0 for model weights, code, and documentation. Teams should still read the licence text in the repositories before building a commercial product on the weights. The release also claims day-0 hardware adaptation covering AMD, Arm, Houmo AI, Intel, and NVIDIA, with further support through the FlagOS stack for chips from vendors including Huawei Ascend, Kunlunxin, MetaX, and Moore Threads, plus Apple Mac on Arm — reported here as the vendors' claim. No on-device performance figures were in the release or on the cards checked, and none are claimed here. Moonshot's Kimi K3.1 Pro and Mistral's Mistral Large 4 are far larger general-purpose models, linked as open-weight peers only; no performance comparison is made or implied.

Illustration created for INORYVA — a quality check on recognised text, with a pass-and-review decision, not a screenshot of MiniCPM-OCR output.

Key Takeaways

• Released 31 October 2026: MiniCPM-OCR 1.0 (1.3B) parses documents; MiniCPM-OCR Jev (2B) inspects OCR output. Jev is not an OCR model — it scores candidates and routes them to auto-acceptance, human review, or best-of-N selection.
• MiniCPM-OCR 1.0 handles scanned, photographed, curved, tilted, and screen-captured documents, dense bilingual Chinese-English content, formulas, and multi-level or borderless tables, with key information extraction for invoices, IDs, and financial statements.
• The release contains no benchmark scores and no licence type. "Top-tier parsing accuracy" and "leading level across mainstream industry evaluations" are vendor wording, reported as such.
• Hugging Face figures, checked 1 November 2026: 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2 for 1.0, at 1.2 pages/s under stated RTX 5090 conditions; Jev at 93.6 AUROC (×100) on the vendors' JevOCRBench, 91.9–93.4 across four parsers. None independently reproduced.
• Licence: Apache License 2.0, confirmed on both cards. Available on GitHub, Hugging Face, and ModelScope, with vLLM, SGLang, llama.cpp, and Ollama support. No download counts, adoption figures, or on-device performance figures verified; none claimed.

Conclusion

Read this release as a workflow, not a leaderboard entry. One small model reads the page, a second checks the reading, and a human looks only at what the checker doubts. If that loop holds on real documents, it removes the most expensive step in large-scale digitisation. The vendors' figures give a reason to test it. Run the models on the ugliest documents in the archive and let the error count, not the announcement, decide whether the inspector earns its place.

Sources

• ModelBest / PRNewswire — "ModelBest and OpenBMB Open-Source MiniCPM-OCR Series, Delivering High-Performance Document Parsing and AI-Powered Quality Control" (31 Oct 2026, 15:24 UTC; two models, Jev's inspector role, vendor wording, availability, runtimes, hardware adaptation; no scores or licence in release text): https://www.prnewswire.com/news-releases/modelbest-and-openbmb-open-source-minicpm-ocr-series-delivering-high-performance-document-parsing-and-ai-powered-quality-control-302925594.html
• Hugging Face — openbmb/MiniCPM-OCR-1.0 model card (93.68 Real5-OmniDocBench, 54.7 OCRBench v2, 1.2 pages/s test conditions, DSpark figures, pipeline, Apache License 2.0; checked 1 Nov 2026 ~03:07 PKT): https://huggingface.co/openbmb/MiniCPM-OCR-1.0
• Hugging Face — openbmb/MiniCPM-OCR-Jev model card (0–1 scoring, 93.6 AUROC on JevOCRBench, 91.9–93.4 cross-parser range, architecture, Apache License 2.0; checked 1 Nov 2026 ~03:07 PKT): https://huggingface.co/openbmb/MiniCPM-OCR-Jev