MiniCPM-V 4.6

Updated
24.08.2026
Vision
Run

MiniCPM-V 4.6 is a 1.3B vision-language model from OpenBMB that reads images and video on-device, with vendor GGUF builds starting at 0.5 GB.

At a glance

  • License: Apache 2.0, free for commercial use
  • Parameters: 1.3B, built on SigLIP2-400M and Qwen3.5-0.8B
  • Context length: 262,144 tokens
  • Modalities: single-image, multi-image and video input
  • Minimum hardware: GGUF from 0.5 GB plus a 1.11 GB vision projector; runs on phones

What is MiniCPM-V 4.6?

MiniCPM-V 4.6 is a 1.3B vision-language model from OpenBMB and the most edge-deployment-friendly model in the MiniCPM-V family to date. It is built on SigLIP2-400M and the Qwen3.5-0.8B LLM, inherits the family's single-image, multi-image and video understanding, and introduces mixed 4x/16x visual token compression so you can trade accuracy against speed at run time. OpenBMB positions it plainly: a pocket-sized MLLM for image and video understanding on your phone.

SpecificationMiniCPM-V 4.6
Total parameters1.3B
ArchitectureSigLIP2-400M vision encoder + Qwen3.5-0.8B LLM
Context window262,144 tokens
ModalitiesText, single-image, multi-image and video input
Token compressionMixed 4x/16x visual token compression
Mobile platformsiOS, Android, HarmonyOS
Release dateApril 2026
LicenseApache 2.0

What MiniCPM-V 4.6 is good at

The pitch is efficiency, and OpenBMB backs it with numbers from its model card. The model scores 13 on the Artificial Analysis Intelligence Index against Qwen3.5-0.8B's 10 with 19x fewer token cost, and beats the larger Ministral 3 3B. On vision-language work it reaches Qwen3.5 2B-level capability on OpenCompass, RefCOCO, HallusionBench, MUIRBench and OCRBench. The LLaVA-UHD v4 technique cuts visual encoding FLOPs by more than half, and token throughput lands around 1.5x that of Qwen3.5-0.8B.

Deployment coverage is unusually broad for a vision model: vLLM, SGLang, llama.cpp and Ollama for inference, SWIFT and LLaMA-Factory for fine-tuning, and quantized variants across GGUF, BNB, AWQ and GPTQ. Every mobile adaptation code path is open-sourced, covering iOS, Android and HarmonyOS.

MiniCPM-V 4.6 hardware requirements

The system requirement to check is memory, and here it is almost nothing. These are OpenBMB's own GGUF builds:

BuildFile size
Q4_00.50 GB
Q4_K_M0.53 GB
Q8_00.81 GB
F161.52 GB
Vision projector (mmproj, F16)1.11 GB

Even the full-precision build with the projector stays under 3 GB, so any recent laptop, a phone, or a Raspberry-class board with enough RAM is a valid target. If the GGUF format is new to you, start with what GGUF is.

How to run MiniCPM-V 4.6 in Atomic Chat

Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.

  1. Download Atomic Chat for your platform and open it.
  2. Search for MiniCPM-V 4.6 in the model browser and open Download Options.
  3. Pick a build, attach an image in the chat, and ask about it.

For what else fits tiny hardware, see the full catalog of local models.

MiniCPM-V 4.6 license

MiniCPM-V 4.6 is released under Apache 2.0. That permits commercial use, modification and redistribution with no royalties, so you can build products on it and run it on your own hardware without a usage fee.

Get the weights from Hugging Face

huggingface-cli download openbmb/MiniCPM-V-4.6
# vendor GGUF build for llama.cpp:
huggingface-cli download openbmb/MiniCPM-V-4.6-gguf MiniCPM-V-4_6-Q4_K_M.gguf mmproj-model-f16.gguf
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniCPM-V-4.6",
    "messages": [{"role": "user", "content": [
      {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
      {"type": "text", "text": "What does this chart show?"}
    ]}]
  }'
from transformers import AutoModel, AutoTokenizer
from PIL import Image

model_id = "openbmb/MiniCPM-V-4.6"
model = AutoModel.from_pretrained(model_id, trust_remote_code=True, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

image = Image.open("receipt.jpg").convert("RGB")
msgs = [{"role": "user", "content": [image, "Extract the total amount."]}]
print(model.chat(msgs=msgs, tokenizer=tokenizer))
import OpenAI from "openai";

const client = new OpenAI({ baseURL: "http://localhost:8000/v1", apiKey: "local" });
const res = await client.chat.completions.create({
  model: "MiniCPM-V-4.6",
  messages: [{ role: "user", content: [
    { type: "image_url", image_url: { url: "https://example.com/screenshot.png" } },
    { type: "text", text: "Describe this screenshot." }
  ]}],
});
console.log(res.choices[0].message.content);
Desktop
macOS
(M1 or better)
Download
Windows
(x64)
Download
Linux
(x86_64)
Download

Frequently asked questions

MiniCPM-V 4.6 is a 1.3B vision-language model from OpenBMB, built on SigLIP2-400M and the Qwen3.5-0.8B LLM. It handles single-image, multi-image and video understanding, and is OpenBMB's most edge-deployment-friendly model to date, with mixed 4x/16x visual token compression for switching between accuracy and speed.

Very little. OpenBMB's own GGUF builds start at 0.5 GB for Q4_0 plus a 1.11 GB vision projector, and the full F16 build is 1.52 GB. That fits comfortably on any recent laptop, and the model is deployable on iOS, Android and HarmonyOS phones with open-sourced adaptation code.

Yes. The weights are released under the Apache 2.0 license, which permits commercial use, modification and redistribution with no royalties. Running it locally has no API fees or per-token charges.

Efficient image and video understanding on small hardware. OpenBMB reports it outperforms Qwen3.5-0.8B on most vision-language tasks and reaches Qwen3.5 2B-level capability on benchmarks including OpenCompass, RefCOCO, HallusionBench, MUIRBench and OCRBench, while its visual encoding uses over 50% fewer FLOPs thanks to the LLaVA-UHD v4 technique.

The model is adapted to vLLM, SGLang, llama.cpp and Ollama for inference, and to SWIFT and LLaMA-Factory for fine-tuning. OpenBMB publishes quantized variants across GGUF, BNB, AWQ and GPTQ formats, so it runs on consumer-grade GPUs and CPUs.