MiniCPM-V 4.6

MiniCPM-V 4.6

Reads single images, several images and video and answers in text, with 1.3B parameters sized for phones and other edge devices. For developers who want a small vision-language 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: MiniCPM-V 4.6 is a small vision-language model from openbmb with 1.3B parameters. It takes images or video plus a question and answers in text. Under the hood it pairs the SigLIP2-400M vision encoder with the Qwen3.5-0.8B language model, and the aim is to run on phones and similar devices.

Who it suits: It suits developers and researchers who want image and video understanding in a small model that can run on a phone or a consumer GPU.

Industry | Who uses it | How they use it | Example
Mobile apps | Developers building on-device AI features | Deploy it on iOS, Android or HarmonyOS with the open-sourced adaptation code | An app team adds a photo-reading feature that runs on the phone
AI research | Researchers testing small vision-language models | Compare its benchmark scores with other small models | A researcher checks it against Qwen3.5-0.8B on OCRBench
Applied machine learning | Teams customizing a model for their own domain | Fine-tune it with SWIFT or LLaMA-Factory on a consumer GPU | A team fine-tunes it for its own kind of images

Anyone who wants the largest vision model regardless of size can skip it, because this one is built to stay small and fast.

Where to use it: It targets phones first, and the files can also be downloaded on a computer.

• iOS, Android and HarmonyOS phones: on-device deployment with the card's adaptation code
• Windows, macOS or Linux: download the files with Python and the Hugging Face tools
• vLLM, SGLang, llama.cpp and Ollama: inference frameworks the model is adapted to
• SWIFT and LLaMA-Factory: fine-tuning ecosystems it supports

Use case | Who | What they get back
Describe or question one image | App builders | A text answer about the picture
Compare several images | Researchers | A text answer that draws on all the images
Understand a video | Developers | A text answer about what happens in the clip

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

• Python and the huggingface_hub package for the download
• A framework from the list above, such as vLLM, llama.cpp or Ollama, to run the model
• About 2.6 GB of disk for the BF16 weights, or about 650 MB at 4-bit as an estimate

1. Install the Hugging Face hub package.
2. Download the model files.
3. Load the model in one of the frameworks the card names.

pip install -U huggingface_hub
hf download openbmb/MiniCPM-V-4.6

How to use it: You give the model an image or a video and ask a question in plain words. The card lists the frameworks it works with and says quantized variants exist in GGUF, BNB, AWQ and GPTQ formats.

Tool | What it does
vLLM and SGLang | Inference frameworks the card says the model is adapted to
llama.cpp and Ollama | Inference frameworks for running it locally, also named in the card
SWIFT and LLaMA-Factory | Fine-tuning ecosystems for customizing it to new domains and tasks
GGUF, BNB, AWQ, GPTQ | Quantized formats in which the card provides variants

Example: You send a photo of a handwritten note and ask, "What does the text in this picture say?" The model answers in text. The card lists OCRBench among the benchmarks where it reaches the level of Qwen3.5 2B.

For newcomers and experts: This is a small AI that looks at pictures or video and replies in words. Your first small step is to download the files with the command above, then try one image with a framework you already know.

The visual side uses SigLIP2-400M, and a technique from LLaVA-UHD v4 cuts visual encoding computation by more than 50%, with about 1.5x the token throughput of Qwen3.5-0.8B. It supports mixed 4x and 16x visual token compression, which lets you trade accuracy for speed.

An image or video goes in
SigLIP2-400M encodes what it sees
Visual tokens are compressed 4x or 16x
The Qwen3.5-0.8B model writes the answer

Good to know:

• The Hugging Face record lists 31 quantized and 21 fine-tuned versions made by other people.
• The benchmark claims are the card's own, and the card also compares token cost and throughput against Qwen3.5-0.8B.
• The card reports that it surpasses the larger Ministral 3 3B on the Artificial Analysis Intelligence Index.

Title: Artificial Analysis Intelligence Index score
MiniCPM-V 4.6: 13 points
Qwen3.5-0.8B: 10 points
Qwen3.5-0.8B-Thinking: 11 points
Ministral 3 3B: 11 points

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: none named for the weights. The card also announces a hosted API service with a public free API key.
• Data it reaches: it downloads files from Hugging Face and reads the images and video you give it.
• Not for: none named in the card.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• 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: 187,584
• Likes: 1,253
• License: apache-2.0
• Quantized versions: 31
• Fine-tunes: 21
• Tags: transformers, safetensors, minicpmv4_6, image-text-to-text, minicpm-v, multimodal, lightweight, conversational

Links

• Model: https://huggingface.co/openbmb/MiniCPM-V-4.6
• Files: https://huggingface.co/openbmb/MiniCPM-V-4.6/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openbmb/MiniCPM-V-4.6

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