MiniCPM-OCR 1.0 GGUF

MiniCPM-OCR 1.0 GGUF

Reads English and Chinese documents, including PDFs, into text, tables and formulas, in GGUF files for llama.cpp. For people who run MiniCPM-OCR 1.0 on their own machine.

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 GGUF holds the llama.cpp ready files of MiniCPM-OCR 1.0, a 1.3B document reader from openbmb. The model turns documents into text, tables, formulas and charts, and can pull out key information. It handles PDFs and documents in Chinese or English. The record lists openbmb/MiniCPM-V-4.6 as its base.

Who it suits: It suits people who want document OCR running on their own computer through llama.cpp, with a choice of file sizes. The card sends anyone after the original weights or benchmarks to the base repository.

Industry | Who uses it | How they use it | Example
Office work | Admin and records staff | Turn scanned pages and PDFs into editable text | A scanned contract page converted into text
Finance | Analysts and accountants | Pull tables and formulas out of reports | A table from a financial report extracted as data
Research | Students and researchers | Read formulas and charts from papers | A page of formulas from a paper parsed into text
Software | Developers who self-host OCR | Run a small quantized reader through llama.cpp | A local service built on one of the GGUF tiers

You can skip it if you want the original weights or benchmark tables, which live in the base repository.

Where to use it:

• llama.cpp, on Windows, macOS or Linux
• The minicpm-ocr Python SDK, which runs the two stage document parsing
• Other self-hosting stacks the project names for MiniCPM-OCR: vLLM, SGLang and Ollama
• The project's free online playground and API, and an iOS app that runs on the device

Use case | Who | What they get back
Parse a document | Office staff | Text, tables and formulas from the page
Extract key information | Analysts | The key details pulled from a document
Run OCR offline | Developers | A reader running on your own machine

How to set it up:

• Python and pip, for the file download
• llama.cpp, since this build is made for it
• One LLM GGUF file plus the vision projector file, because every tier needs the projector
• PaddlePaddle or PaddleX if you use the SDK's layout stage

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Choose one LLM GGUF tier and keep the projector file with it.
4. Start llama.cpp with the LLM file in model and the projector in mmproj.

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

How to use it: Two flags tell llama.cpp which files to load. The card also says to prefer the Q8_0 tier for pages heavy with code or XML.

Tool | What it does
--model | Takes the LLM GGUF file for the tier you picked
--mmproj | Takes the vision projector file, which every tier needs
llama.cpp | Runs the GGUF build
minicpm-ocr SDK | Splits a document into regions and reads each one
PP-DocLayoutV3 | Detects document regions and their reading order

The next table shows what each tier is for.

Tier | What it is for
F16 | The reference build with no quantization
Q8_0 | The card's recommended balance, nearly identical to F16 on tables and dense text
Q6_K | A middle tier with good quality for its size
Q5_K_M | A smaller tier that keeps strong quality
Q4_K_M | The smallest tier the card recommends
Q4_0 | The smallest file in the list

Example: You load the Q8_0 file with the projector and give it a scanned report page that holds a table. You ask for the table's contents, and the text of the table comes back.

For newcomers and experts: This is a reader for scanned pages and PDFs. For a first try, download the files, pick the Q8_0 tier, load it in llama.cpp with the projector and give it one clean page.

The projector downsamples by 4 times, and no 16 times variant exists for this model. The card says Q4_K_M stays accurate on tables, loses a little on dense text, and may drop closing tag slashes in code heavy documents. Full parsing uses a two stage pipeline from the SDK. A layout model finds regions first, and the stages can run in different environments and on different machines.

A page goes to the SDK
PP-DocLayoutV3 finds regions and reading order
The SDK crops each region
The GGUF model reads each region

Good to know: Every tier needs the projector, so budget for it on top of the tier you choose. File sizes in GB:

Title: File size in GB
Projector, required: 0.901 GB
F16: 1.516 GB
Q8_0: 0.812 GB
Q6_K: 0.630 GB
Q5_K_M: 0.578 GB
Q4_K_M: 0.529 GB
Q4_0: 0.501 GB

• License: Apache 2.0
• Cost to run: no key or paid service is named for these files, and the usage guide offers a free online playground and API for the main model
• Data it reaches: downloads files from Hugging Face, and the guide says self-hosting keeps data on your side
• Not for: a 16 times projector, which does not exist for this model, or the original weights and benchmarks, which are in the base repository

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: minicpm-ocr
• Parameters: 752M
• Weights: about 376 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 58,162
• Likes: 133
• License: apache-2.0
• Languages: en, zh
• Based on: openbmb/MiniCPM-V-4.6
• Tags: minicpm-ocr, gguf, document-parsing, llama.cpp, multimodal, quantized, image-text-to-text, conversational

Links

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

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