MiniCPM-V-4.7-1B-gguf

MiniCPM-V-4.7-1B-gguf

Runs a 1B vision-language model through llama.cpp, answering text prompts about images. GGUF files in many quant sizes, for developers and hobbyists who want a small local model.

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: This repository holds GGUF quantized files of MiniCPM-V 4.7 1B, a vision-language model of about 1B parameters that answers text prompts about images through llama.cpp. The record estimates about 503 MB at 4-bit. The unquantized weights and a fuller card live with the original MiniCPM-V-4.7-1B model from openbmb.

Who it suits: It suits developers and hobbyists who want a small image-understanding model on a laptop, a phone or another device with limited resources.

Industry | Who uses it | How they use it | Example
Mobile apps | App developers | Build image questions into an iOS, Android or HarmonyOS app | Describe a photo a user just took
Local AI | Hobbyists | Chat about pictures on a laptop through the Llama app | Ask what is in a screenshot
Edge hardware | Embedded engineers | Run a small model on chips the card names | Test image captions on a device with limited memory

Anyone who needs the original BF16 weights should use the original repository instead.

Where to use it: You can run it in a desktop app, in a terminal or on a device.

• The Llama desktop app on macOS and Windows, made jointly by the llama.cpp developers and Hugging Face
• A terminal, with llama cli for chat and llama serve for a local web page and API
• iOS, Android and HarmonyOS, which the card says MiniCPM-V 4.7 supports
• Chips from Rockchip, Houmo, Intel, AMD and Arm, where the card says it has day-0 support

Use case | Who | What they get back
Describe an image | Anyone with a laptop | A written description of the picture you pass in
Local image API | Developers | An OpenAI-compatible endpoint served from their machine
On-device vision | Mobile and edge teams | A small model that fits limited hardware

How to set it up: The one-click route needs only the Llama app. The terminal route needs llama installed.

• macOS or Windows for the one-click route, or a terminal with llama installed
• Python, only if you download the files by hand
• Free disk space for the quant you pick

1. Install the Llama app. On macOS use the first command below, and on Windows use its installer.
2. Open the model's Hugging Face page, press the Use this model button and pick Llama. The app picks a build suited to your hardware, fetches it and starts a chat window.
3. Or skip the app and use the terminal commands under How to use it.
4. To fetch the files by hand, install the hub tools and download the repository.

brew install --cask llama-app

pip install -U huggingface_hub
hf download openbmb/MiniCPM-V-4.7-1B-gguf

How to use it: In a terminal, the -hf flag pulls the model and its matching mmproj file for you.

llama cli   -hf openbmb/MiniCPM-V-4.7-1B-gguf
llama serve -hf openbmb/MiniCPM-V-4.7-1B-gguf     # WebUI + OpenAI-compatible API on :9931

If you already have the files on disk, point -m at one quant and add the common mmproj file. The second command turns thinking on and asks for more tokens with -n.

# default (thinking disabled, no flag needed)

# thinking enabled
llama cli -m thinking/MiniCPM-V-4_7-1B-Q4_K_M-thinking.gguf --mmproj mmproj-model-f16.gguf \
          --image img.jpg -p "Describe." -n 1024

Put your own value in its key setting first.

Two sets of files exist. They differ only in the chat template built into them, and their weights are identical. Each set spans quant types from F16 down to IQ3.

File set | Thinking | When to pick it
Repository root | Off, and it cannot be switched on | The recommended default, since the model was not trained for reasoning
thinking folder | On by default | You want to read the reasoning steps, and you can afford more tokens

Tool | What it does
llama cli | Runs the model in a terminal chat, with an image if you pass one
llama serve | Starts a web page and an OpenAI-compatible API on port 9931
llama-mtmd-cli and llama-server | Older entry points that stay available for development
Llama app | One-click desktop app that downloads a fitting build

Example: You run llama cli with the Q4_K_M file, the mmproj file, an image called img.jpg and the prompt "Describe." The model prints a short description of the picture. With the thinking files it also prints a reasoning trace before the answer.

For newcomers and experts: If you are new, install the Llama app and pick this model from its Hugging Face page. That opens a chat window where you can drop in a picture and ask about it.

For someone who knows the field, the card says the model compresses visual tokens early inside the vision encoder, following a technique from LLaVA-UHD v4, and supports mixed 4x and 16x token compression. That lets you trade visual detail against cost. The card also says it beats Qwen3.5 0.8B on overall score and efficiency, but this page gives no numbers. The runtime adds enable_thinking to the template by itself. As a result, the --reasoning off flag does nothing for the root set, while for the thinking set it hides the trace.

You pass an image and a prompt
The mmproj file encodes the image
The 1B language model reads both
It prints an answer, or a trace first

Good to know: The mmproj files are shared by both sets because they carry no chat template. Thinking multiplies the token count, so raise -n on the thinking files.

• License: Apache-2.0
• Cost to run: none named
• Data it reaches: the -hf flag fetches the model and mmproj from Hugging Face. llama serve opens a web page and API on port 9931 of your own machine.
• Not for: reasoning tasks on the root files, where thinking is off and cannot be turned on

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: llama.cpp
• Parameters: 1B
• Weights: about 503 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 168,636
• Likes: 114
• License: apache-2.0
• Based on: openbmb/MiniCPM-V-4.7-1B
• Fine-tunes: 1
• Tags: llama.cpp, gguf, quantized, imatrix, dspark, speculative-decoding, minicpm-v, multimodal, lightweight, image-text-to-text, conversational

Links

• Model: https://huggingface.co/openbmb/MiniCPM-V-4.7-1B-gguf
• Files: https://huggingface.co/openbmb/MiniCPM-V-4.7-1B-gguf/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openbmb/MiniCPM-V-4.7-1B-gguf

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