Overview
MiniCPM-OCR 1.0 is a document recognition suite OpenBMB open-sourced on October 27, 2026 under Apache 2.0, with code and weights released together. It contains two models: MiniCPM-OCR 1.0 at 1.3 billion parameters handles parsing, and MiniCPM-OCR Jev at 2 billion parameters handles quality control.
The split is the distinctive part. Most OCR returns one block of text and leaves you to find the mistakes. Jev scores each region of a parse from 0 to 1, evaluating text, formulas and tables separately, so only low-scoring regions reach a human review queue. Jev is not tied to its sibling: OpenBMB measured it over results from MiniCPM-OCR, GLM-OCR, PaddleOCR-VL-1.6 and Xiaomi OCR-0, with AUROC between 91.9 and 93.4.
On the parsing side the scores are 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2. DSpark decoding lifts end-to-end efficiency by 55 percent, putting throughput at roughly 1.49 times PaddleOCR-VL-1.6 and 3.62 times MinerU2.5 Pro, about 1.2 pages per second on an RTX 5090. Day-one support covers vLLM, SGLang, llama.cpp and Ollama, with BF16, GGUF, AWQ and GPTQ weights.
Key Features
- Parser and grader kept separate: The 1.3B model reads the page and the 2B Jev scores the result independently. The grader can be swapped out or pointed at another parser's output.
- Region-level confidence: Jev returns a 0 to 1 score for text, formulas and tables separately, routing questionable regions to review. OpenBMB publishes review-free curves for legal documents, financial receipts and complex documents, plotting accuracy against review skipped.
- Real capture conditions: Training covers scans, phone photos, curved book pages, skewed pages and recaptured screens. Chinese and English are supported, with output as Markdown or JSON covering tables, formulas, charts, seals and key information extraction.
- DSpark decoding: With DSpark-BF16 and DSpark-GGUF decoding, end-to-end efficiency rises 55 percent and throughput reaches about 1.49 times PaddleOCR-VL-1.6 and 3.62 times MinerU2.5 Pro.
- Four runtimes on day one: vLLM, SGLang, llama.cpp and Ollama all work at release, with BF16, GGUF, AWQ and GPTQ weights plus multi-chip builds, a pip SDK and a full API.
- Still fine-tunable: The model is trained for OCR on top of MiniCPM-V 4.6, pairing a SigLIP vision encoder with a Qwen3.5 language model, and can be trained further on your own data.
Use Cases
- Bulk entry of contracts, invoices and financial reports, with Jev's scores concentrating human review on a small share of regions.
- Digitising phone photos and recaptured archives, since curved and skewed pages are in the training distribution.
- Building the document front door for a knowledge base, where Markdown output drops straight into retrieval.
- Local or edge deployment, where GGUF and Ollama builds run on machines without a GPU.
- Auditing another OCR system by attaching Jev to its output and scoring the results.
- Key information extraction that pulls fields from forms rather than transcribing everything.
Pros
- Parsing and grading are separate models, so only low-confidence regions reach a reviewer.
- Four runtimes supported at launch, which broadens the deployment options.
- Weights ship in BF16, GGUF, AWQ and GPTQ, covering a wide hardware range.
- DSpark decoding lifts end-to-end efficiency 55 percent, with throughput ahead of comparable open-source parsers.
- Apache 2.0 across code, docs and weights.
- Jev works with third-party parsers, so it has value outside this suite.
Pricing
Apache 2.0 across code, documentation and weights, with no extra commercial restriction. Self-hosting carries only hardware cost, and OpenBMB measures about 1.2 pages per second on an RTX 5090, with real throughput depending on backend, concurrency and page content. Online demos are available for both China and global users to test on real documents, and any hosted service bills according to its own platform.
Summary
The value here is admitting that OCR gets things wrong and marking where. Most teams do not lose to a percentage point of accuracy, they lose because nobody knows which cell on which page needs a human, so they either review everything or wait for a downstream failure. Jev turns that into a score plus a threshold that fits into an existing pipeline. Parsing speed and runtime support are adequate, and a 1.3B footprint makes local deployment realistic.
Version History
- MiniCPM-OCR 1.0 and the Jev quality model open sourced (2026-10-27): OpenBMB released MiniCPM-OCR 1.0, a 1.3B parser, and MiniCPM-OCR Jev, a 2B quality judge, under Apache 2.0. Parsing scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2, with DSpark decoding adding 55 percent end-to-end efficiency and throughput about 1.49 times PaddleOCR-VL-1.6 and 3.62 times MinerU2.5 Pro. Jev returns 0 to 1 confidence per region for text, formulas and tables, with AUROC between 91.9 and 93.4 across parsers. Day-one support for vLLM, SGLang, llama.cpp and Ollama, with BF16, GGUF, AWQ and GPTQ weights.