BuildTube Preview About
Back to the feed
On-device MoE inference speed comparison, with optimization on the left and without optimization on the right

Picture published by the creator on Hugging Face.

MiniCPM-V-4.7-35B-A3B: Ask about images and video on phones

Added · Updated

Made by OpenBMB

View profile on Hugging Face (opens in a new tab)

Ask questions about one image, several images or a video and get text answers.

For developers who want image and video understanding on phones and local devices.

Model details

Original model openbmb/MiniCPM-V-4.7-35B-A3B

Understand images/documentsWrite & reason

Input Image or text Output Text

Try the browser demo (opens in a new tab)Follow the model card setup (opens in a new tab)Use through the OpenBMB API (opens in a new tab)

Ways to use it Browser demo · Run locally · Use through an API

Local fit guide

One reviewed local configuration. Add My computer details to compare them.

BF16 vLLM on NVIDIA · vLLM · The pinned card lists vLLM as an inference framework and provides its Transformers inference examples. This estimate uses the BF16 checkpoint and a 4K comparison context.

How this configuration was checked

96 GB VRAM · reviewed estimate Source lists 70.4 GB (file-size). Estimate uses 70.43 GB exact BF16 checkpoint bytes, 0.084 GB 4K KV cache (10 full-attention layers of 40 hybrid layers × 2 KV heads × 256 head width × K/V × 2-byte values), 4 GB runtime buffers, and 21.486 GB remaining in the 96 GB VRAM tier.

BF16 · 4,096 token context · source context 262,144 tokens · batch size 1 · concurrency 1 · Single-user 4K comparison on the source vLLM path; the model also accepts visual inputs.

Source checked 2026-10-21 · revision 896c2d454672476a7b45dcc5e3f0c700a0218577. pinned card lists vLLM and BF16 Transformers inference; exact checkpoint shard bytes and hybrid attention config supply the estimate inputs

Set up My computer · Open the exact setup source (opens in a new tab)
Creator exampleSee the source example (opens in a new tab)

What it does

MiniCPM-V 4.7 is a multimodal model, meaning an AI that understands more than text. The source says it builds on the family's single-image, multi-image and video understanding. This variant is a mixture-of-experts model, where only part of the model is used for each piece of text. The creators say it compresses visual information early inside its vision encoder, which reduces computing, and that developers can choose a mixed 4x/16x compression to trade visual detail against cost.

How to use it

Input → action → output: give it an image, several images or a video plus your question, run it, and read the model's answer. Get the weights from Hugging Face and run them with a framework the source lists: SGLang, vLLM, llama.cpp or Ollama. The page also lists compressed (quantized) versions in GGUF, BNB, AWQ and GPTQ formats, and a live demo link under Get started.

Common use cases

  • Show the model one image and ask what is in it.
  • Give it several images together and ask a question about them.
  • Give it a video and ask it to answer questions about what happens.

Can I use this?

Needs a model runtime on your computer

Requirements The model is open source under Apache 2.0 and not gated, so no access request is needed. Download the weights from the model page and use a supported framework. The creators say it supports iOS, Android and HarmonyOS. Exact RAM or GPU needs for the full model are not documented.

Caveat The creators report decoding at 13+ tokens per second with 3 GB of memory on an iPhone 17 Pro Max, using their own llama.cpp-omni technique and quantization. That is their claim for one device. The source also notes that mixture-of-experts storage needs are not limited to the active parameters.

Worth knowing The speed figure is the creators' claim for one iPhone; results on other devices are not documented. Licensed under Apache License 2.0.

Why it's here

Not trending this week

Last seen on the Hugging Face trending list on 19 October 2026. It stays on BuildTube so you can still find it.

  • 1,115 people have liked it on Hugging Face.
  • It was downloaded 637,750 times in the last 30 days.

Numbers checked on 8 November 2026; not refreshed since.

Behind it

View model page

BuildTube has not run or verified this project. Everything above is written from what the creator published.

Made this?

Quick questions

What is MiniCPM-V-4.7-35B-A3B?
MiniCPM-V 4.7 is a multimodal model, meaning an AI that understands more than text.
What can I use MiniCPM-V-4.7-35B-A3B for?
Show the model one image and ask what is in it; Give it several images together and ask a question about them; Give it a video and ask it to answer questions about what happens
Can I run MiniCPM-V-4.7-35B-A3B on my computer?
It runs on your own hardware with a model runtime. The creators say it supports iOS, Android and HarmonyOS. Exact RAM or GPU needs for the full model are not documented.
What do I need to use MiniCPM-V-4.7-35B-A3B?
Open source under Apache 2.0 and not gated. Needs a supported framework such as SGLang, vLLM, llama.cpp or Ollama.
What license does MiniCPM-V-4.7-35B-A3B use?
Licensed under Apache License 2.0.

More from this week

Related projects

Explore by topic

Post this

A ready-made picture of this card for X, LinkedIn or a group chat. Link to this page alongside it.

Share image for MiniCPM-V-4.7-35B-A3BDownload share image

For makers

Made this? Show it’s on BuildTube — add the badge to your README or embed the card on your site.

Featured on BuildTube

README badge (Markdown)

[![Featured on BuildTube](https://buildtube.app/badge/hf-openbmb-minicpm-v-4-7-35b-a3b.svg)](https://buildtube.app/p/hf-openbmb-minicpm-v-4-7-35b-a3b/)

Embed card (HTML)

<iframe src="https://buildtube.app/embed/hf-openbmb-minicpm-v-4-7-35b-a3b/" width="560" height="240" style="border:0;max-width:100%" loading="lazy" title="MiniCPM-V-4.7-35B-A3B on BuildTube"></iframe>

Back to the feed