MiniCPM-V-4.7-1B

Name: MiniCPM-V-4.7-1B
Author: openbmb

by openbmb

The 1.5B-parameter MiniCPM-V 4.7: images, several images and video in, text out, small enough for a phone

340likes

1.55Bparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

openbmb/MiniCPM-V-4.7-1B

Feature URI

Deploy MiniCPM-V-4.7-1B

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

Read the customer-hosted guide Talk to us about single-tenant

Overview

MiniCPM-V 4.7 1B is the small member of OpenBMB's MiniCPM-V 4.7 family, released on 6 October 2026 under Apache 2.0. It reads one image, several images or a video and answers in text, so it can caption a frame, read the text in a screenshot or describe what happens in a clip. The weights hold about 1.55B parameters: a 27-layer vision encoder in front of a 24-layer Qwen3.5 language model.

OpenBMB pitches it on capability for its size. The card says it improves on MiniCPM-V 4.6 and beats Qwen3.5 0.8B on overall score and inference efficiency in Instruct mode, with the figures shown as a chart rather than a table. Its card runs it on an iPhone 17 Pro Max and lists iOS, Android and HarmonyOS support plus day-0 support for Rockchip, Houmo, Intel, AMD and Arm chips.

For search, it is a cheap captioner: run it over every keyframe or image to get text you can embed and filter on. The 16x token compression mode is the fast setting and 4x keeps finer detail. We have not measured it.

Architecture

A 27-layer vision encoder (hidden size 1152, 14-pixel patches, 980-pixel base image size) compresses visual tokens early inside the encoder, using the intra-ViT compression from LLaVA-UHD v4, and feeds a 24-layer Qwen3.5 text model (hidden size 1024, 8 attention heads, 262,144-token context). About 1.55B parameters in bfloat16. downsample_mode chooses 16x or 4x visual-token compression per request; images are sliced into tiles (max_slice_nums, 36 recommended for images and 1 for video), and video is sampled at up to max_num_frames frames (128 in the card's example).

Mixpeek SDK Integration

# Caption each keyframe with MiniCPM-V-4.7-1B, then store the caption as metadata
# so Mixpeek can embed it and filter on it alongside the frame.
import requests

for frame in frames:  # [{"url": "s3://bucket/video_017/frame_0042.jpg", "caption": "..."}]
    requests.post(
        "https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
        headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
        json={
            "blobs": [{"property": "image", "type": "image", "url": frame["url"]}],
            "metadata": {"caption": frame["caption"]},
        },
    )

Capabilities

Answers questions about single images, multiple images and video
Mixed 16x or 4x visual-token compression, chosen per request
Runs on iOS, Android and HarmonyOS, and through vLLM, SGLang, llama.cpp and Ollama
GGUF, BNB, AWQ, GPTQ and FP8 quantized variants; Apache 2.0

Use Cases on Mixpeek

Captioning every keyframe of a video library so the captions can be embedded and searched

Reading on-screen text in screenshots and slides on the device that captured them

Describing images in a photo library without sending them to a hosted API

A small first-pass model that decides which frames deserve a larger model

Performance

Input SizeImages (sliced into up to max_slice_nums tiles), several images, or video at up to max_num_frames frames; 262,144-token context

Embedding Dimn/a (generates text)

GPU LatencyInput dependent

GPU ThroughputBatch dependent

GPU MemoryModel dependent

OpenBMB reports its results against MiniCPM-V 4.6 and Qwen3.5 0.8B as charts, without a table to quote. Its card shows it running on an iPhone 17 Pro Max. We have not measured it.

Common Pipeline Companions

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

The larger mixture-of-experts sibling, for frames where the 1B's answer is not good enough

google/embeddinggemma-2

Embeds the captions, and the images themselves, into one space for search

Frequently Asked Questions

What is MiniCPM-V 4.7 1B?

A 1.55B-parameter vision-language model from OpenBMB, released on 6 October 2026 under Apache 2.0. It reads images, several images or a video and answers in text, and it is built to run on phones and laptops.

Can MiniCPM-V 4.7 1B run on a phone?

Yes. OpenBMB's card shows unedited recordings of it running on an iPhone 17 Pro Max, and lists iOS, Android and HarmonyOS support along with GGUF and other quantized variants.

What is the difference between 16x and 4x downsample_mode?

16x merges visual tokens for speed and is the default. 4x keeps four times as many visual tokens, which helps with small text and fine detail at a higher compute cost. The value has to be passed to both the processor and generate().

How does it compare with MiniCPM-V 4.7 35B-A3B?

The 35B-A3B is a mixture-of-experts model with about 3B parameters active per token and a 40-layer language model. The 1B is a dense 24-layer model, much cheaper to run and less capable, which suits a first pass over many frames.

How do I use MiniCPM-V 4.7 1B with Mixpeek search?

Caption each image or keyframe with it, store the caption in the object's metadata, and let Mixpeek embed and index it next to the image, as in the example on this page. A retriever can then search the captions and the pixels together.