openbmb/ minicpm-v4.7

6,367 Downloads Updated 4 weeks ago

MiniCPM-V 4.7 — the latest multimodal LLM in the MiniCPM-V series, in 1B and 35B-A3B (MoE) variants. Single-image, multi-image, and video understanding, with smooth on-device MoE inference.

vision 1b

ollama run openbmb/minicpm-v4.7

curl http://localhost:11434/api/chat \
  -d '{
    "model": "openbmb/minicpm-v4.7",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='openbmb/minicpm-v4.7',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'openbmb/minicpm-v4.7',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

10 models

Size / Usage

Context

Input

minicpm-v4.7:latest

22GB · 256K context window · Text, Image · 4 weeks ago

minicpm-v4.7:latest

22GB

256K

Text, Image

minicpm-v4.7:1b

1.8GB · 256K context window · Text, Image · 4 weeks ago

minicpm-v4.7:1b

1.8GB

256K

Text, Image

Readme

MiniCPM-V 4.7

GitHub · Hugging Face · ModelScope · Cookbook · Demo

MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, released in two variants — 1B (dense) and 35B-A3B (MoE, the first MoE model in the series). It builds on the family’s single-image, multi-image, and video understanding capabilities, combining stronger foundation capabilities with efficient visual encoding and on-device inference.

Highlights

🔥 A New Pareto Frontier in Performance and Efficiency. The 1B variant substantially outperforms Qwen3.5 0.8B in both overall score and inference efficiency; the 35B-A3B variant achieves a higher overall score in less inference time than Qwen3.6 35B-A3B.
📱 Smooth On-Device MoE Inference. Combining Lookahead Expert Scheduling (LES), fine-grained layer-wise quantization, and DSpark speculative sampling, the 35B-A3B variant runs at 13+ tokens/s on iPhone 17 Pro Max — without compromising capability.
🚀 Efficient Architecture. Intra-ViT early visual compression (from LLaVA-UHD v4) with mixed 4x/16x visual token compression, to trade off detail against inference cost.
💻 Broad Platform Support. iOS, Android, and HarmonyOS, with day-0 support for Rockchip, Houmo, Intel, AMD, and Arm chip platforms.
🛠️ Developer Friendly. Supported by SGLang, vLLM, llama.cpp, and Ollama; fine-tuning via SWIFT and LLaMA-Factory; GGUF / BNB / AWQ / GPTQ / FP8 quantized variants available.

Available tags

Tag	Quantization	Size	Notes
`1b`	Q4_K_M	1.8 GB	recommended 1B default
`1b-q4_K_M`	Q4_K_M	1.8 GB
`1b-q8_0`	Q8_0	2.2 GB	higher quality
`1b-fp16`	F16	3.1 GB	full precision
`35b-a3b`	Q4_K_M	22 GB	recommended 35B-A3B default
`35b-a3b-q4_K_M`	Q4_K_M	22 GB
`35b-a3b-q8_0`	Q8_0	38 GB	higher quality
`35b-a3b-fp16`	F16	70 GB	full precision
`latest`	Q4_K_M	22 GB	same as `35b-a3b`

Quick start

”`bash ollama run openbmb/minicpm-v4.7

# MiniCPM-V 4.7

<p align="center">
  <a href="https://github.com/OpenBMB/MiniCPM-o">GitHub</a> &nbsp;·&nbsp;
  <a href="https://huggingface.co/openbmb/MiniCPM-V-4.7-1B">Hugging Face</a> &nbsp;·&nbsp;
  <a href="https://modelscope.cn/models/OpenBMB/MiniCPM-V-4.7-1B">ModelScope</a> &nbsp;·&nbsp;
  <a href="https://github.com/OpenSQZ/MiniCPM-V-CookBook">Cookbook</a> &nbsp;·&nbsp;
  <a href="https://huggingface.co/spaces/openbmb/MiniCPM-V-4.7-Demo">Demo</a>
</p>

**MiniCPM-V 4.7** is the latest model in the MiniCPM-V series, released in two variants — **1B** (dense) and **35B-A3B** (MoE, the first MoE model in the series). It builds on the family's single-image, multi-image, and video understanding capabilities, combining stronger foundation capabilities with efficient visual encoding and on-device inference.

## Highlights

- 🔥 **A New Pareto Frontier in Performance and Efficiency.** The **1B** variant substantially outperforms Qwen3.5 0.8B in both overall score and inference efficiency; the **35B-A3B** variant achieves a higher overall score in less inference time than Qwen3.6 35B-A3B.
- 📱 **Smooth On-Device MoE Inference.** Combining [Lookahead Expert Scheduling (LES)](https://github.com/tc-mb/llama.cpp-omni/blob/master/docs/les/lookahead-expert-scheduling.md), fine-grained layer-wise quantization, and DSpark speculative sampling, the 35B-A3B variant runs at **13+ tokens/s on iPhone 17 Pro Max** — without compromising capability.
- 🚀 **Efficient Architecture.** Intra-ViT early visual compression (from [LLaVA-UHD v4](https://github.com/THUMAI-Lab/LLaVA-UHD-v4)) with **mixed 4x/16x visual token compression**, to trade off detail against inference cost.
- 💻 **Broad Platform Support.** iOS, Android, and HarmonyOS, with day-0 support for Rockchip, Houmo, Intel, AMD, and Arm chip platforms.
- 🛠️ **Developer Friendly.** Supported by SGLang, vLLM, llama.cpp, and Ollama; fine-tuning via SWIFT and LLaMA-Factory; GGUF / BNB / AWQ / GPTQ / FP8 quantized variants available.

## Available tags

| Tag | Quantization | Size | Notes |
|:---|:---|:---|:---|
| `1b` | Q4_K_M | 1.8 GB | recommended 1B default |
| `1b-q4_K_M` | Q4_K_M | 1.8 GB | |
| `1b-q8_0` | Q8_0 | 2.2 GB | higher quality |
| `1b-fp16` | F16 | 3.1 GB | full precision |
| `35b-a3b` | Q4_K_M | 22 GB | recommended 35B-A3B default |
| `35b-a3b-q4_K_M` | Q4_K_M | 22 GB | |
| `35b-a3b-q8_0` | Q8_0 | 38 GB | higher quality |
| `35b-a3b-fp16` | F16 | 70 GB | full precision |
| `latest` | Q4_K_M | 22 GB | same as `35b-a3b` |

## Quick start

```bash
ollama run openbmb/minicpm-v4.7

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)