OpenBMB releases MiniCPM-OCR 1.0, a 1.3B OCR model scoring 93.68 on Real5-OmniDocBench

OpenBMB has released MiniCPM-OCR 1.0, a 1.3B-parameter model that scores 93.68 on Real5-OmniDocBench, a benchmark for real-world document parsing. It is the first OCR-specific model in the MiniCPM-V/o family, according to the Hugging Face model card. Weights, code and documentation are released under the Apache License 2.0.

What the model does

MiniCPM-OCR 1.0 targets text-centric document understanding: text, tables, formulas, charts, seals, and key information extraction (KIE). Its primary languages are Chinese and English. OpenBMB says it handles scans, curved pages, photographs, skewed pages and recaptured documents, as well as PDFs.

Benchmarks and throughput

All figures below are reported by OpenBMB.

• Real5-OmniDocBench: 93.68
• OCRBench v2: 54.7
• Other benchmarks: OpenBMB claims leading results on OmniDocBench v1.6, Wild-OmniDocBench and olmOCR-Bench, but the card text does not give those scores.
• Comparison set: Among the open-source OCR models compared, OpenBMB says MiniCPM-OCR 1.0 has the highest equal-weight average across OmniDocBench v1.6, Real5-OmniDocBench, Wild-OmniDocBench and olmOCR-Bench.
• Throughput: On an NVIDIA GeForce RTX 5090, running all 1651 OmniDocBench v1.6 images at image-level concurrency 1, the model reaches 1.2 pages/s. OpenBMB puts this at about 1.49× PaddleOCR-VL-1.6 and 3.62× MinerU2.5 Pro, with each system in its accelerated configuration.
• DSpark: With DSpark decoding acceleration (BF16, GGUF), OpenBMB says end-to-end processing efficiency improves by 55%.

OpenBMB notes that real throughput depends on page content, backend, concurrency and hardware, and that scores from different benchmarks are not directly comparable.

Architecture

The model is trained for OCR on top of MiniCPM-V 4.6. It combines a SigLIP vision encoder with a Qwen3.5 language model. Visual tokens use 4× compression, with scale_resolution 448 and up to 9 slices.

Full-document parsing uses a pipeline: the minicpmvocr SDK reads PDF, PNG or JPG input, PP-DocLayoutV3 detects regions and reading order, and the model recognises each cropped region concurrently. Results are assembled into Markdown in reading order. Tables become HTML and formulas stay LaTeX. The prompt must match the region type (text, table, formula, seal, or chart to table).

For KIE, users send an image with a field-extraction prompt and receive a JSON object, with no page Markdown step required. OpenBMB advises checking the model-generated JSON before using it downstream.

Deployment

The card lists support from day one for vLLM, SGLang, llama.cpp and Ollama, plus Transformers and FlagOS (multi-chip). The vLLM example sets --max-model-len 40960. Released checkpoints:

• BF16 (original)
• GGUF
• AWQ W4A16
• GPTQ W4A16
• DSpark draft models in BF16 and GGUF

DSpark requires a matching main and draft model from the same release. Fine-tuning is documented for LLaMA-Factory and ms-swift, and OpenBMB also ships a pip SDK, a full API and agent workflows. The Hugging Face page shows 791 downloads in the last month and says no Inference Provider currently deploys the model. Pricing is not applicable to the open weights, and hosted API pricing is not disclosed.

What this means

MiniCPM-OCR 1.0 is a 1.3B model for document parsing that comes with an Apache 2.0 license, quantized builds and support for common local and server runtimes. That gives teams working with Chinese and English documents a model they can self-host and fine-tune. The benchmark and throughput comparisons come from OpenBMB, and the card points to a reproduction guide for the test settings. The throughput figure was measured on a single RTX 5090 at concurrency 1, so results on other hardware or under heavier concurrency may differ.