MiniCPM-OCR 1.0

MiniCPM-OCR 1.0

A 1.3B model from OpenBMB that reads documents in English and Chinese, parsing text, tables, formulas, charts and seals and pulling out key fields.

Made by openbmb on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MiniCPM-OCR 1.0 is a 1.3B model from OpenBMB for reading documents. Among the MiniCPM-V and MiniCPM-o models, it is the first one made only for OCR.

Besides PDFs, it works on pages captured in everyday conditions: scans, curved or skewed pages, photos and re-photographed documents. It takes an image in and gives text out.

Who it suits: It suits developers who turn documents into structured text and care about speed. Teams that cannot send files to an outside service can host it on their own machines.

Industry | Who uses it | How they use it | Example
Finance and accounting | Operations teams | Read invoices and statements into fields | Pulling a total from a photographed invoice
Legal | Document teams | Turn scanned contracts into searchable text | Parsing a scanned agreement
Education and research | Researchers | Convert papers with formulas and tables | Extracting a table from a PDF
Software | Developers | Add document parsing to an app or agent | An upload form that reads an ID page

Skip it if your documents are in languages other than English and Chinese, since the record lists only those two.

Where to use it:

• vLLM, SGLang, llama.cpp and Ollama for self-hosting, with an official Docker image
• The model's own pip SDK and its API
• An online playground where you can try it for free
• An iOS demo app that runs on the device, where document parsing works on iOS only for now
• Windows, macOS or Linux for the file download

Use case | Who | What they get back
Document parsing | Developers | Text, tables and formulas from a page image or PDF
Key information extraction | Automation teams | Fields pulled out for downstream systems
Chart and seal reading | Analysts | The text and data found in charts and seals

How to set it up: You need a few things in place first.

• Python and pip, for the download command
• A serving tool such as vLLM, SGLang, llama.cpp or Ollama, or the pip SDK
• About 2.6 GB of disk for the BF16 weights, from the record

1. Install the Hugging Face hub package with pip.
2. Download the repository with the hf command.
3. Serve it with one of the listed tools, or use the SDK. The model card and the project's deployment guide hold the steps.

pip install -U huggingface_hub
hf download openbmb/MiniCPM-OCR-1.0

How to use it: You send the model an image of a page and ask it to parse the document or extract fields. You can do this through a self-hosted server, through the SDK or through the hosted API. The card also offers an agent skill file, which lets an agent install the model and parse a first document from one prompt.

Example: You photograph a printed invoice and send it to your self-hosted model with a request to extract the key fields. The model returns the text and the key information, such as the vendor and the total, ready for your system.

For newcomers and experts: The quickest first try is the online playground, which needs no setup. Once you like the output, move to a local server so your documents stay on your side.

The card reports 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2. With DSpark decoding acceleration it says end-to-end efficiency improves by 55 percent, and throughput reaches about 1.49 times PaddleOCR-VL-1.6 and 3.62 times MinerU2.5 Pro. Among the open models it compared, it reaches 1.2 pages per second.

Title: Throughput of MiniCPM-OCR 1.0 against other models, from the card
Against PaddleOCR-VL-1.6: 1.49 times
Against MinerU2.5 Pro: 3.62 times

Send a page image or a PDF
The model reads text and layout
It parses tables, formulas and charts
You get text and key fields back

Good to know:

• License: Apache 2.0
• Cost to run: the card offers a free API to try it, and self-hosting needs your own hardware
• Data it reaches: self-hosting keeps documents on your side, and the hosted API sends them to an online service
• Not for: languages other than English and Chinese, since the record lists only those two

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: minicpm-ocr
• Parameters: 1.3B
• Weights: about 2.6 GB in BF16; about 650 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 128,483
• Likes: 457
• License: apache-2.0
• Languages: en, zh
• Based on: openbmb/MiniCPM-V-4.6
• Quantized versions: 9
• Fine-tunes: 9
• Tags: minicpm-ocr, safetensors, minicpm_ocr, document-parsing, multimodal, vision-language, table-recognition, formula-recognition, image-text-to-text, conversational

Links

• Model: https://huggingface.co/openbmb/MiniCPM-OCR-1.0
• Files: https://huggingface.co/openbmb/MiniCPM-OCR-1.0/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openbmb/MiniCPM-OCR-1.0

Then follow the usage in its model card. Nothing in this kit runs by itself.