MiniCPM-V 4

MiniCPM-V Multimodal Series

MiniCPM-V 4 is a 4.1B-parameter multimodal model built on a SigLIP2-400M + MiniCPM4-3B architecture.

Key Features

• OpenCompass average score of 69.0, surpassing GPT-4.1-mini and MiniCPM-V 2.6 (8.1B)
• Efficient on-device: first-token latency <2s and generation speed >17 tokens/s on iPhone 16 Pro Max
• Outperforms Qwen2.5-VL-3B, which has a comparable parameter count

Model Download

Version	HuggingFace	ModelScope
MiniCPM-V 4.0	Download	Download

Quick Start

Python

import torch
from transformers import AutoModel, AutoTokenizer
from PIL import Image

model = AutoModel.from_pretrained(
    "openbmb/MiniCPM-V-4",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16
).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained("openbmb/MiniCPM-V-4", trust_remote_code=True)

image = Image.open("image.jpg").convert("RGB")
msgs = [{"role": "user", "content": [image, "描述图片内容。"]}]
answer = model.chat(msgs=msgs, tokenizer=tokenizer)
print(answer)

Fine-tuning

Bash

cd MiniCPM-V
MODEL="openbmb/MiniCPM-V-4"
LLM_TYPE="qwen"
sh finetune/finetune_lora.sh

Related Resources

Topic	Link
Fine-tuning	MiniCPM-V 4 fine-tuning guide
Code repository	OpenBMB/MiniCPM-V
Official model documentation	minicpm_v4_en.md
On-device apps and downloads	MiniCPM-V-Apps downloads