Skip to content
Find local AI

Your computer: not set · Set it

Writing

MiniCPM-V 4.7 35B-A3B

Apache 2.0

OpenBMB's larger vision model for pictures, documents, and video. The 35B-A3B size is a mixture of experts.

PhotosPhoto

Compare two models

looks at pictures, documents, and video

The 3 GB figure is for an iPhone 17 Pro Max using Lookahead Expert Scheduling, which loads experts from storage as they are needed. The full weights still have to be stored. On a desktop with ordinary llama.cpp or Ollama, plan for the whole Q4_K_M file plus the vision projector, about 20.8 GB.

19.7 GB · · needs ~20.8 GB · smallest file

Runs in:

Fits on

Desktops

Macs

Mini PCs & servers

Flavors:File · 19.7 GB

Tasks this model can do

How good is it?

OpenBMB's larger vision model for pictures, documents, and video. The 35B-A3B size is a mixture of experts.

This job is not on the July 2026 chat board, so the line above comes from the model card.

Clean printed pages come out well. Handwriting, messy photos, and tiny type are weaker. Tesseract and PaddleOCR are light. olmOCR, Qwen2.5-VL, and MiniCPM-V 4.7 read the page more like a person. The larger MiniCPM-V file needs about 21 GB while it runs.

Will it fit?

A green light means it fits and feels fine. Amber means it fits, with a wait or a tight memory pool. Red means the memory it needs is bigger than the room we count.

HONEST COMPARISON

Local or cloud: Describe what's in a photo?

Reviewed Oct 2026

  • Privacy: Local winsThe photo stays on your computer
  • Cost: TieFree options exist on both sides
  • Quality: Cloud winsFewer invented details on hard scenes
  • Ease: Cloud winsDrop the image in a chat

Running MiniCPM-V 4.7 35B-A3B locally

A local vision model can say what is in a picture without uploading it.

  • Private photos stay on disk
  • Works offline
  • You can ask a follow-up in the same chat

Where it falls short

  • It misses small text and rare objects
  • It can invent a detail that is not there
See local tools for this task ↓

Cloud & apps

The largest cloud vision models are more reliable on cluttered scenes and tiny text.

  • CChatGPTCloud AIFreemiumPopular choice

    Describes a photo and answers a follow-up in the same thread.

    Best for: Cluttered scenesFollow-up questions

    The photo is uploaded

    Visit ChatGPT (opens in new tab)
  • CClaudeCloud AIFreemium

    Careful descriptions when you need the wording to stay literal.

    Best for: Literal captions

    The photo is uploaded

    Visit Claude (opens in new tab)
  • GGeminiCloud AIFreemium

    Strong on photos already in a Google account.

    Best for: Google photosQuick looks

    The photo is sent to Google

    Visit Gemini (opens in new tab)

Our take

Describe private photos locally. Use a cloud vision model when the scene is cluttered and the picture can be shared.

Choose local if…

  • The photo is private
  • You are offline

Choose cloud if…

  • The scene is cluttered
  • You need the small text read

Will it run on my computer?

Pick the memory you have. A Mac counts 75% of unified memory. A graphics card counts 90% of its memory. A PC without a graphics card counts half of its RAM. The memory we plan around is 20.8 GB.

MiniCPM-V 4.7 35B-A3B needs more room. It needs 20.8 GB, and this setup leaves about 12 GB.

Official install guides

We link to the makers. We do not write our own install steps.