MiniCPM-V 4.7 1B has a small quantized language model, but its local vision installation contains a second substantial file. OpenBMB’s Q4_K_M language weights and default vision projector total approximately 1.78 GB of downloads. Ollama packages the pair; llama.cpp requires you to select both files explicitly.

The official 1B checkpoint supports image, multiple-image and video understanding under Apache 2.0. This guide concerns that dense model. The larger variant has a separate expert-scheduling deployment story, whose device results should not become performance expectations for this installation.

Ollama offers the shortest starting point

The OpenBMB Ollama listing supplies the exact tag:

ollama run openbmb/minicpm-v4.7:1b

Following the Ollama multimodal CLI convention, provide an image using its local path alongside your question. Begin with one legible document and one narrow request, such as reading its invoice number. Verify the returned characters against the original before expanding the task.

The registry currently lists this package at 1.8 GB, including a 672 MB Q4_K_M language component and a 1.1 GB projector. The download figure describes packaged weights. Runtime memory also depends on context, visual processing, intermediate buffers and the chosen backend. It cannot establish a laptop’s minimum RAM by itself.

The projector outweighs the quantized language file

The official GGUF file inventory makes the storage calculation reproducible. The language file contains 672,003,680 bytes; mmproj-model-f16.gguf contains 1,110,117,984 bytes. Adding them gives 1,782,121,664 bytes, or 1.782 GB using decimal units.

Quantizing the language component further leaves the projector untouched. That limits how much the entire installation can shrink from language quantization alone. Budget both components before choosing a download, and distinguish disk capacity from resident memory.

The repository also provides mmproj-model-f16-4x.gguf, at 901,033,888 bytes. Paired with the same Q4_K_M language file, it totals approximately 1.573 GB. That smaller projector file belongs to a different visual-compression configuration; its size does not imply less downstream visual-token work.

llama.cpp needs matching, current-format files

OpenBMB’s llama.cpp deployment instructions warn that older conversion scripts can produce incorrect M-RoPE computation or a missing projector tensor. Version 4.7 adds rope.section_order metadata and the v.tok_embd_sep tensor. Use official 4.7 GGUF files and a current compatible runtime.

After building llama.cpp using its linked CMake instructions, download both files into its models directory. With the Hugging Face CLI installed, run these commands from the llama.cpp checkout:

hf download openbmb/MiniCPM-V-4.7-1B-gguf \
  MiniCPM-V-4_7-1B-Q4_K_M.gguf --local-dir ./models
hf download openbmb/MiniCPM-V-4.7-1B-gguf \
  mmproj-model-f16.gguf --local-dir ./models

./build/bin/llama-mtmd-cli \
  -m ./models/MiniCPM-V-4_7-1B-Q4_K_M.gguf \
  --mmproj ./models/mmproj-model-f16.gguf \
  -c 8192 --reasoning off --image ./document.jpg \
  -p "Read the invoice number and total exactly as printed."

Replace document.jpg with your own image. The recipe adapts the documented invocation; EyesTech has not executed it on a test device. Explicit --reasoning off selects a plain reply because the template enables thinking under the automatic setting. The 8,192-token context is a configuration choice, not a memory guarantee.

Compression deserves an accuracy check

The model’s Transformers usage guidance describes 16x visual compression and a 4x setting that retains four times as many visual tokens for finer detail. In Transformers, pass the same downsample_mode to both apply_chat_template and generate; otherwise placeholder construction and vision processing can disagree. In GGUF deployment, use the corresponding projector variant.

For small print, evaluate both settings on identical documents. Count exact field matches, missing fields and wrong digits. Record complete response time, including image processing, and peak memory after a cold start. Those observations would establish whether retaining more detail justifies the extra work on your machine.

The useful starting point is one model, its matching projector and a verified document task. A successful load establishes compatibility. Correct extraction at acceptable latency establishes usefulness.

Prepared with AI assistance from primary documentation and file-size calculations. Cover: AI-generated conceptual illustration. No independent device benchmarks are claimed.

Categorized in:

Architectural Teardowns, Benchmarks & Hype Checks, Local LLMs & Homelabs,

Last Update: October 14, 2026