MiniCPM-OCR-1.0

Name: MiniCPM-OCR-1.0
Author: openbmb

by openbmb

OpenBMB's 1.3B-parameter OCR model for document parsing: text, tables, formulas, charts, seals and field extraction from scans and photos, under Apache 2.0

369likes

1.3Bparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

openbmb/MiniCPM-OCR-1.0

Feature URI

Deploy MiniCPM-OCR-1.0

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

Read the customer-hosted guide Talk to us about single-tenant

Overview

MiniCPM-OCR 1.0 is a 1.3B-parameter OCR model from OpenBMB, released in October 2026 and the first OCR-specific model in the MiniCPM-V family. It is trained for OCR on top of MiniCPM-V 4.6 and reads text, tables, formulas, charts, seals and structured fields from Chinese and English documents and PDFs, including scans, curved, skewed and photographed pages. Paired with the PP-DocLayoutV3 layout model through its SDK, it turns a page into Markdown. Weights, code and documentation are released under Apache 2.0.

Architecture

A SigLIP vision encoder joined to a Qwen3.5 language model, 1.3B parameters in total, fine-tuned for OCR from MiniCPM-V 4.6. Visual tokens are compressed 4x, with a scale_resolution of 448 and up to 9 slices per image. Full-page parsing runs as a pipeline: PP-DocLayoutV3 finds the regions, MiniCPM-OCR reads each one with a task prompt, and the minicpmvocr SDK assembles the result. An optional DSpark draft model speeds decoding through speculative decoding.

Mixpeek SDK Integration

# Parse a PDF to Markdown with MiniCPM-OCR 1.0 (commands from the OpenBMB tutorial).
python -m pip install --upgrade "minicpm-ocr[all]"

# 1. Serve the model with vLLM (add the tutorial's --speculative-config block for DSpark).
vllm serve openbmb/MiniCPM-OCR-1.0 \
  --served-model-name minicpm-ocr \
  --host 127.0.0.1 --port 18499 \
  --max-model-len 40960 \
  --trust-request-mm-kwargs \
  --default-chat-template-kwargs '{"enable_thinking":false}'

# 2. Parse: the SDK runs PP-DocLayoutV3 for layout and sends each region to the model.
minicpmvocr parse \
  --device cuda:0 \
  --server-url http://127.0.0.1:18499 \
  --model minicpm-ocr \
  --input doc.pdf \
  --output doc.md

Capabilities

Text recognition for paragraphs, titles and whole pages
Tables and charts returned as HTML tables, formulas as LaTeX
Seal recognition and key information extraction to JSON
Chinese and English documents, including scans, curved pages and photos
Serves through vLLM, SGLang, llama.cpp and Ollama, with GGUF, AWQ and GPTQ checkpoints

Use Cases on Mixpeek

Turning scanned PDFs and phone photos of documents into Markdown before chunking them for RAG

Pulling named fields (invoice number, date, total) out of receipts and forms as JSON

Reading tables and charts in financial statements into rows you can store and query

A small OCR model you can run on one GPU next to the files instead of calling a hosted API

Benchmarks

Dataset	Metric	Score	Source
Real5-OmniDocBench	Overall score	93.68	https://huggingface.co/openbmb/MiniCPM-OCR-1.0
OCRBench v2	Score	54.7	https://huggingface.co/openbmb/MiniCPM-OCR-1.0

Performance

Input SizeDocument images and PDF pages

GPU LatencyInput dependent

GPU Throughput1.2 pages/s on one RTX 5090 over the 1,651 OmniDocBench v1.6 images, accelerated configuration (reported by OpenBMB)

GPU MemoryModel dependent

We have not measured it. The throughput figure is OpenBMB's, at image-level concurrency 1; it depends on page content, backend and hardware.

Common Pipeline Companions

PaddlePaddle/PP-DocLayoutV3

The layout model its SDK uses to find text, table and formula regions

PaddlePaddle/PaddleOCR-VL-1.6

Another compact document parser; OpenBMB compares throughput against it

baidu/Unlimited-OCR-2

A larger OCR model built for long documents

Frequently Asked Questions

What is MiniCPM-OCR 1.0?

A 1.3B-parameter OCR model from OpenBMB, fine-tuned from MiniCPM-V 4.6 for document parsing. It reads text, tables, formulas, charts, seals and named fields from Chinese and English documents, including scans and photographed pages.

What does MiniCPM-OCR output for tables and formulas?

Tables and charts come back as HTML tables after the SDK's post-processing, formulas as LaTeX, and plain text for paragraphs and seals. For key information extraction you list the fields in the prompt and get a JSON object back, with "###" for a field that is not on the page.

Can I use MiniCPM-OCR commercially?

Yes. OpenBMB releases the weights, code and documentation under the Apache License 2.0.

How fast is MiniCPM-OCR?

OpenBMB reports 1.2 pages per second on one NVIDIA RTX 5090 across the 1,651 OmniDocBench v1.6 images, with DSpark decoding enabled. It also reports about 1.49 times the throughput of PaddleOCR-VL-1.6. We have not measured it.

How do I use it with Mixpeek?

Run it where your documents are, store each page's Markdown and extracted fields with the file and page number, and upsert them into a collection so the text is searchable and the fields filterable. To search the page images directly, without an OCR step, use a multimodal embedding model in the same collection.