MiniCPM-OCR Jev GGUF

MiniCPM-OCR Jev GGUF

Scores OCR readings of text, formula and table regions with a quality value between 0 and 1, in GGUF files for llama.cpp. For people who check OCR results 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 Jev GGUF holds the llama.cpp ready files of MiniCPM-OCR Jev, a 2B model from openbmb. Jev does not read a page itself. It grades OCR output for three kinds of regions, text, formulas and tables, and returns a quality value between 0 and 1.

Who it suits: It suits teams that run OCR and want an automatic check on how good each reading is. The card points anyone after the original weights to the base repository.

Industry | Who uses it | How they use it | Example
Document OCR | Teams building OCR pipelines | Score each candidate reading before keeping it | Comparing two readings of one table region by their scores
Quality control | Review and QA staff | Spot weak OCR results by their low scores | Flagging a formula region that scores low
Research | OCR researchers | Evaluate candidate outputs on text, formula and table regions | Scoring several readings of the same region
Software | Developers who self-host models | Run a local scoring server | One llama-server instance on a single machine

You can skip it if you want a model that reads documents itself, since this one scores readings.

Where to use it:

• llama.cpp and its llama-server, on Windows, macOS or Linux
• Your own computer, with Python for the file download
• Other formats from the same project: the original BF16, AWQ W4A16 and GPTQ W4A16

Use case | Who | What they get back
Score an OCR reading | OCR developers | A quality value from 0 to 1
Filter weak output | QA staff | A way to separate low and high scoring readings
Self-host a scorer | Engineers | A local server that answers scoring requests

How to set it up:

• Python and pip, for the file download
• llama.cpp with its llama-server program
• One LLM GGUF file and the vision projector file, which is required
• Disk space for the tier you choose, from about 1 GB to about 3.8 GB

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Start llama-server with your chosen tier and the projector.

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

The card's server command uses the Q8_0 tier:

llama-server \
  --model ./MiniCPM-OCR-Jev-Q8_0.gguf \
  --mmproj ./mmproj-model-f16.gguf \
  --host 0.0.0.0 --port 18498 \
  --ctx-size 40960 \
  --n-gpu-layers 999

How to use it: Start the server with the command above, then send it scoring requests. Check the verdict field in each response, because only a verdict of scored means scoring took place.

Tool | What it does
llama-server | Starts a local server for the model
--model | Takes the LLM GGUF file for your tier
--mmproj | Takes the vision projector file
--port 18498 | Sets the port the server listens on
--ctx-size 40960 | Sets the context size
--n-gpu-layers 999 | Asks for the model layers to be placed on the GPU
verdict | A response field that says whether scoring ran

The tiers differ in size. F16 is the unquantized reference, Q8_0 is near lossless, and Q4_K_M is the smallest the card recommends. The four IQ tiers are marked imatrix and are the smallest files.

Example: You pass the server an image region and an OCR reading of it, such as a table. The response gives a quality value from 0 to 1 and a verdict, and you trust the value only when the verdict says scored.

For newcomers and experts: This model grades OCR output rather than producing it. For a first try, download the Q8_0 tier with the projector, start the server with the command above and look at the verdict in a response.

Underneath, it is a 2B model with a separate vision projector, served by llama.cpp with a 40960 token context in the card's sample. The same project also publishes the original BF16 weights and two 4-bit builds, AWQ W4A16 and GPTQ W4A16, for other runtimes.

An OCR reading and its region arrive
llama-server loads the GGUF and projector
The model scores the reading
A value from 0 to 1 and a verdict return

Good to know: The project keeps a FAQ and a llama.cpp guide for Jev. The FAQ covers setup, model files, memory needs and odd results. File sizes in GB:

Title: File size in GB
Projector, required: 0.671 GB
F16: 3.776 GB
Q8_0: 2.012 GB
Q6_K: 1.556 GB
Q5_K_M: 1.411 GB
Q4_K_M: 1.274 GB
Q4_0: 1.202 GB
IQ3_XS: 1.027 GB

• License: Apache 2.0, covering weights, code and documentation
• Cost to run: no key or paid service is named, you supply the hardware
• Data it reaches: downloads files from Hugging Face, and the sample command binds the server to all network interfaces with --host 0.0.0.0, so check your firewall on a shared network
• Not for: reading documents on its own, or the original weights and full card, which are in the base repository

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: llama.cpp
• Parameters: 1.9B
• Weights: about 941 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 31,316
• Likes: 22
• License: apache-2.0
• Languages: en, zh
• Based on: openbmb/MiniCPM-OCR-Jev
• Tags: llama.cpp, gguf, minicpm-ocr, quality-assessment, multimodal, quantized, image-text-to-text, decision-model, feature-extraction

Links

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

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