MiniCPM-V-4.7-35B-A3B
by openbmb
OpenBMB's first mixture-of-experts MiniCPM-V: 35B parameters, about 3B active per token, for images, multi-image and video, built to run on a phone
openbmb/MiniCPM-V-4.7-35B-A3BDeploy MiniCPM-V-4.7-35B-A3B
Single-tenantMixpeek 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.
Overview
MiniCPM-V 4.7 35B-A3B is a vision-language model from OpenBMB, released in October 2026 under the Apache 2.0 license. It reads single images, several images and video, and answers in text, so it can caption, read text in an image, parse a PDF page, ground an object or describe what happens in a clip. It is the first mixture-of-experts model in the MiniCPM-V series: 35B parameters in total, with 8 of 256 experts routed per token.\n\nThe card's focus is running it on a device. It pairs the model with lookahead expert scheduling, which preloads the experts the next layer is likely to need, plus layer-wise quantization and a speculative-decoding draft model, and reports more than 13 tokens a second on an iPhone 17 Pro Max. Visual tokens are compressed inside the vision encoder, and a downsample_mode setting trades detail (4x) against speed (16x).
Architecture
A 27-layer vision encoder (hidden size 1152, 14-pixel patches, 980-pixel base image size) with early visual-token compression, feeding a 40-layer Qwen3.5-MoE language model (hidden size 2048, 256 experts with 8 routed per token plus a shared expert, 262,144-token context, one multi-token-prediction layer). About 35B parameters in total and about 3B active per token. Images are sliced into tiles (max_slice_nums); video is sampled at up to 128 main frames, optionally with extra sub-frames per second stacked into a grid image.
Mixpeek SDK Integration
// MiniCPM-V 4.7 runs on your side. Store what it writes about each asset as
// payload text next to an embedding of that text, so the collection can search it.
const res = await fetch(
"https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
{
method: "POST",
headers: {
Authorization: "Bearer API_KEY",
"Content-Type": "application/json",
},
body: JSON.stringify({
collection_id: "col_your_collection",
documents: [
{
document_id: "scene-00412",
payload: { scene_description: modelOutput, source_key: "archive/2026/clip-00412.mp4", start: 84.0 },
vectors: { "multimodal-embedding": embeddingOfModelOutput },
},
],
}),
},
);
// Managed alternative, if this exact model is not the requirement:
// universal_extractor@v1 runs google/gemini-embedding-2
// (3072-d) over a bucket, with no inference of your own.Capabilities
- Answers questions about single images, multiple images and video
- Reads text in images and parses PDF pages, per the card's cookbook
- Visual grounding of objects described in text
- 4x or 16x visual-token downsampling, chosen per request
- Runs in Transformers 5.18+, SGLang, vLLM, llama.cpp and Ollama
- Apache 2.0 license
Use Cases on Mixpeek
Performance
The card publishes its benchmark scores as charts, comparing against Qwen3.6 35B-A3B in Instruct mode, so no figures are copied here. On-device speed per the card: more than 13 tokens a second on an iPhone 17 Pro Max. We have not measured it.
Common Pipeline Companions
Frequently Asked Questions
What is MiniCPM-V 4.7?
An open vision-language model from OpenBMB that answers questions about images, multiple images and video. The 35B-A3B version is a mixture-of-experts model with about 3B parameters active per token, released under Apache 2.0.
Can MiniCPM-V 4.7 run on a phone?
That is what the release is built for. The card reports more than 13 tokens a second on an iPhone 17 Pro Max using lookahead expert scheduling, layer-wise quantization and speculative decoding, and lists iOS, Android and HarmonyOS support.
How does MiniCPM-V 4.7 handle video?
It samples up to 128 main frames by default and can add sub-frames for each second, stacked into one grid image, so short actions are not missed. The card recommends 3 or 5 sample points per second and one slice per frame for video.
What does downsample_mode do?
It sets how many visual tokens each image becomes. 16x, the default, merges tokens for speed; 4x keeps four times as many for fine detail such as small text. It has to be passed to both apply_chat_template and generate.
Specification
Research Paper
MiniCPM-V 4.7 model card
arxiv.orgBuild a pipeline with MiniCPM-V-4.7-35B-A3B
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free