New SOTA in OCR accuracy and efficiency: ModelBest open-sources MiniCPM-OCR and the first document AI inspector
On October 28, ModelBest and OpenBMB officially released and open-sourced the MiniCPM-OCR series for efficient document parsing and text recognition. The release includes two models: MiniCPM-OCR 1.0 (1.3B) and MiniCPM-OCR Jev.
As the first OCR-dedicated model in the MiniCPM family, MiniCPM-OCR 1.0 delivers SOTA capability and extreme efficiency in a 1.3B on-device lightweight form, pushing the efficiency-performance Pareto frontier of open-source OCR models. It achieves a SOTA score of 93.68 on Real5-OmniDocBench, a document parsing benchmark for real-world scenes. With DSpark decoding acceleration, end-to-end processing efficiency improves by 55%, achieving higher parsing accuracy at lower per-page inference latency.
Overall comparison of open-source OCR models on efficiency and effectiveness: throughput on the x-axis, average score across four benchmarks on the y-axis; MiniCPM-OCR leads on both dimensions
To address the long-standing industry pain point that "parsing still requires manual page-by-page verification," MiniCPM also launches the industry's first OCR parsing-result decision model, MiniCPM-OCR Jev. As a quality-inspection "System-1" model, it offers cross-model general adaptability, automatically judging the parsing quality of text, formulas, and tables, reaching an overall AUROC of 0.936 and greatly reducing the burden of manual review.
This article includes video content. Please view it in the original WeChat post.
No limits on real-world conditions: scans, curved pages, and re-photographed images all parse accurately
Most documents in the real world aren't flat, standardized, high-resolution scans — invoices snapped with a phone, thick contracts naturally warped at the binding, academic papers with long complex formulas, re-photographed images with screen glare and moiré patterns...
Slow parsing, errors on complex layouts, and high manual review costs have become the "three mountains" blocking document digitalization.
Our path is clear: within an extremely lightweight model capacity, teach the model to build robust visual representations under complex physical capture conditions and layout structures. MiniCPM-OCR 1.0 (1.3B) is built for complex, changeable real-world scenes:
No tedious preprocessing; accurate parsing even from casual phone shots: on Real5-OmniDocBench, which specifically examines real-world scenes (covering scans, photos, curved pages, skew, and re-photographed images), MiniCPM-OCR 1.0 achieves a SOTA score of 93.68. Without tedious pre-correction, the model reliably restores layout structure and text content, greatly lowering the barrier to document digitalization.
No collapse on complex layouts; accurate bilingual Chinese-English structured restoration: on OmniDocBench v1.6, Wild-OmniDocBench, olmOCR-Bench, and CHAOS-Bench, which specifically tests extremely chaotic layouts, the model shows leading parsing quality. This means that for dense Chinese-English mixed content, complex math formulas, and multi-level merged-cell or borderless tables in research papers, engineering reports, and cross-language business materials, the model delivers high-quality, clearly structured restoration.
General recognition and information extraction: on the comprehensive OCRBench v2, the model achieves a SOTA score of 54.7. Beyond full-page document restoration, the model also supports key information extraction (KIE), accurately extracting key fields from invoices, certificates, and reports per business instructions, eliminating the tedium of secondary manual transcription.
Six-benchmark radar chart: MiniCPM-OCR 1.0 has the largest overall outline
Six-benchmark bar charts by scenario: MiniCPM-OCR 1.0 is in the top tier across all and ranks first in four
Full comparison of 14 mainstream OCR models across six benchmarks: MiniCPM-OCR 1.0 leads across the board and ranks first in four
DSpark decoding acceleration: pushing the efficiency boundary of large-scale document throughput
In enterprise document processing, beyond accuracy, per-page processing time and compute ownership cost directly determine whether a solution can be deployed. Especially in throughput tasks involving tens of millions of pages — financial contract archiving, academic library digitization, government and enterprise historical archive cleanup — compute costs often become a heavy burden.
MiniCPM-OCR 1.0 introduces DSpark decoding acceleration, improving end-to-end processing efficiency by 55% and reaching 1.6x the original throughput, with shorter processing cycles and lower compute costs.
We ran a relative throughput comparison under unified hardware and data benchmarks (RTX 5090, the full OmniDocBench v1.6 test set, default configurations of each system's official SDK). Measured in pages per second, MiniCPM-OCR 1.0 reaches 1.49x the throughput of PaddleOCR-VL-1.6 and 3.62x that of MinerU2.5 Pro. With the same hardware investment and time window, MiniCPM-OCR 1.0 can process multiples of document pages, truly reducing costs and improving efficiency in scaled business deployment.
Throughput comparison: MiniCPM-OCR 1.0 reaches 1.217 pages/s (all tested models use accelerated configurations)
The industry-first OCR decision model Jev: an "AI inspector" goes live, ending manual page-by-page review
Imagine this: a team processes 10,000 documents with OCR. How do they confirm the results are fine?
Full manual verification is costly and slow; random spot checks can hardly eliminate missed errors in key formulas or core amounts. An OCR model's overall score on a benchmark reflects its average level across many samples. But in serious commercial and production environments such as contracts, finance, and research, every document must be absolutely reliable. This forces enterprises to invest substantial manpower in review, making it the most labor-intensive link in document automation.
MiniCPM's solution: give the pipeline an "AI inspector" — MiniCPM-OCR Jev. MiniCPM-OCR Jev is the industry's first OCR parsing-result decision model. It doesn't parse text; it focuses on judging the reliability of parsing results.
In Thinking, Fast and Slow, psychologist Daniel Kahneman divides human thinking into two kinds: System 1 is fast intuitive reaction, such as recognizing a friend's face at a glance; System 2 is slow step-by-step reasoning, such as solving a complex math problem. MiniCPM-OCR Jev takes the System-1 intuitive decision route: without long reasoning, it compares the original image with the OCR output and quickly makes a confidence judgment on parsing reliability:
Precise multimodal judgment: carefully compares the original image with the output text, sharply catching missing characters, wrong characters, or format misalignment in body text, LaTeX formulas, and structured tables.
Stable cross-model generalization: Jev's judgment is independent of the specific parsing model. It has excellent cross-model generalization and can plug into various existing OCR systems as an independent "inspector," enabling a seamless upgrade without replacing the original parsing model. On parsing results from MiniCPM-OCR, GLM-OCR, PaddleOCR-VL-1.6, and Xiaomi-OCR-0, Jev's AUROC stays stable between 91.9 and 93.4.
Judgment accuracy approaching much larger models: with only 2B parameters, MiniCPM-OCR Jev reaches an overall AUROC of 93.6 on JevOCRBench, an OCR parsing-result decision benchmark. It not only surpasses the dedicated judgment model JEV-35B-A3B (AUROC: 90.1) but also approaches the judgment level of the much larger Qwen3.5-397B-A17B (AUROC: 93.8).
Full JevOCRBench comparison: the 2B MiniCPM-OCR Jev surpasses 35B-class judgment models
With Jev, business processes can change substantially:
Automatic filtering, less manual review: set score thresholds based on business scenarios and accuracy requirements; regions above the threshold can be automatically stored without manual review. In legal document scenarios, an exemption rate as high as 85% achieves 99.4% exempted accuracy; in financial invoices, complex documents, and other scenarios, the exemption scope can be gradually expanded according to the corresponding accuracy requirements, freeing people from aimless "carpet sweeps" to focus on key decisions and greatly reducing verification hours.
Automatic filtering: exemption rate and exempted accuracy curves under different thresholds
Multi-model selection for better parsing quality: based on the same layout segmentation result, multiple OCR models parse each region separately, then Jev scores the candidate results, selects the high-scoring result per region, and combines them into the output, bringing stable accuracy gains to the outputs of multiple mainstream parsing solutions.
Multi-model selection: gains from Jev scoring for different OCR base systems
Real business cases: from handwritten notes to financial reports, parsing and decision-making at full power
Beyond the numbers, look at how the model performs on real documents. MiniCPM-OCR targets industries where documents are the main business carrier — finance, tax, legal, and healthcare — and has been deployed in many customer scenarios.
The model can locally parse and structure documents such as invoices, receipts, contracts, transaction records, and reports: restoring layout and table structures, merging cross-page content, and delivering results as Markdown or field-structured data for business systems to consume. Quality verification of parsing results can be done with the MiniCPM-OCR Jev decision model.
Chinese scenarios: handwritten notes, complex matrix formulas, multi-column newspapers, wide financial tables, official documents with seals, illustrations converted to structured tables, and key information extraction (KIE) under seal occlusion — seven high-frequency scenarios at a glance:
Chinese scenario case set (images): handwritten notes, complex formulas, newspapers, wide financial tables, sealed official documents, illustration-to-table conversion, key information extraction
The Chinese pipeline — from the original image and Layout boxes (seal and text region detection) to structured Markdown rendering:
Chinese scenario case set (GIF): original image — layout detection — structured output, the full flow
English scenarios: handwritten physics notes, yellowed historical documents, scanned official documents, multi-layout newspapers, multi-fiscal-year financial reports, and key information extraction from thermal receipts:
English scenario case set: handwritten notes, historical documents, scanned official documents, newspapers, financial reports, receipt information extraction
The English parsing and output process:
English scenario case set (GIF): original image — layout detection — structured output, the full flow
Now Jev's judgment. For the same original image and two candidate parsing results, Jev's scores align almost perfectly with true quality: in the Chinese scenario, a candidate with multiple semantic errors scores only 0.0067, while the accurate version scores 0.984; formula and table scenarios are equally clear-cut. English scenarios follow the same pattern — candidates with concentrated errors and scrambled structure are pushed below 0.1, while reliable results stay above 0.8. This means business teams can draw a line with the score: auto-store online, and check only the low scores offline.
Jev English judgment cases (images): score comparison across text, formula, and table scenarios
Open-source adaptation across the ecosystem: out of the box, minimal development
We hope these capabilities can be directly used and modified by more developers. The MiniCPM-OCR series continues ModelBest's technical DNA of "open-source, efficient, developer-friendly," completing comprehensive software and hardware adaptation on the first day of release:
Day-0 chip adaptation: leveraging a long-built chip partner ecosystem, MiniCPM-OCR has completed Day-0-level adaptation with AMD, Arm, Houmo AI, Intel, NVIDIA, and others. In addition, the FlagOS community completed Day-0 cross-chip adaptation for MiniCPM-OCR 1.0 and MiniCPM-OCR Jev based on a unified multi-chip AI system software stack, covering 12 platforms: CIX Technology, Huawei Ascend, Hygon, Iluvatar CoreX, Kunlunxin, MetaX, Moore Threads, Xiwang, T-Head, Tsingmicro, NVIDIA, and Mac (ARM). The corresponding multi-chip BF16 versions are now available on ModelScope and Hugging Face; developers can search for FlagRelease to get model weights and deployment instructions. In this round of single-card BF16 evaluation, the FlagOS version of Jev reaches 98.7% of NVIDIA's native solution in compute throughput; on-device, Jev uses only the M5 CPU to complete one page-block judgment in an average of 1.20 seconds. With FlagOS cross-chip adaptation, industry users can choose document parsing and quality-inspection deployment options based on business needs and existing compute.
Currently, ModelBest has established deep collaboration with chip and computing platform companies including AXERA, D-Robotics, Huawei Ascend, Hygon, MediaTek, Qualcomm, SiEngine, and UNISOC. The broad hardware-software ecosystem covers phones, PCs, cars, robots, edge devices, and servers, enabling MiniCPM models to land on all kinds of terminal devices faster and more efficiently.
Inference and fine-tuning framework support: the MiniCPM-OCR series is fully compatible with mainstream inference frameworks such as vLLM, SGLang, llama.cpp, and Ollama; on the client side it supports the Hugging Face CLI, pip SDK, a full-featured API, and a mature Agent Skill plugin wrapper, with fine-tuning code fully open-sourced. A smooth online WebDemo is also live, giving developers a complete path from online experience to efficient device-cloud deployment to domain-specific training.
Want to try it directly? We've launched the official WebDemo — upload a casually taken document or paper and see the parsing results in real time. The model weights, deployment toolchain, and online services for MiniCPM-OCR 1.0 and MiniCPM-OCR Jev are now fully open and available; developers and enterprise partners can get resources and try the services on GitHub, Hugging Face, and ModelScope.
We welcome everyone to download, test, and share feedback, as we explore the application boundaries of next-generation on-device document intelligence together.