A Tiny Vision-Language Model for Fast, High-Volume Image Questions
MiniCPM-V 4.7 1B is OpenBMB's smallest vision-language model. Ask it about one image, several images at once, or plain text, and it answers in text. It compresses images early inside the vision encoder, so it stays fast and light — and you can switch to a finer 4× mode when small details matter. Apache 2.0.
📸 See it in action
The new hire review. One office photo and a question a manager might ask. Default settings (16× compression, greedy decoding).

Prompt
This is the newest employee at our company. What is this employee's job title, and are they doing a good job? Answer as a short, honest performance review.
MiniCPM-V 4.7 1B's answer (unedited)
Title: New Office Dog (or maybe a new team lead?). Job: Probably a "Pet-Approved" or "Canine Coordinator" — but honestly? They're doing a good job. Sitting at the desk, ready to handle papers, with that tie and golden fur, they bring calm and charm. Not your usual human, but they're clearly here to help.
Input image made with Z-Image-Turbo in CNAPS Studio. The answer is the model's real output from one run.
What it does
MiniCPM-V 4.7 1B takes a text prompt plus up to three images and returns a text answer. Despite its size it is a general assistant, not a single-purpose reader: it can describe a photo, answer a question, compare images or pull details out of a screenshot.
Its trick is early visual-token compression inside the vision encoder. By default each image is compressed 16×, which keeps runs quick and cheap. Switch downsample_mode to 4× to keep four times as many visual tokens when you need small text, dense diagrams or fine detail.
OpenBMB says the 1B model clearly outperforms Qwen3.5 0.8B in instruct mode.
Problem it solves
Use case | How it helps |
High-volume tagging and captions | Fast enough for Batch Runs over large image sets |
Simple yes/no checks | "Is there a person in this photo?" before routing to the next node |
Compare up to three images | Connect all three image ports in one call |
Fine detail when needed | Switch to 4× mode for small text or diagrams |
Text-only prompts | Works with no image at all |
Input/Output
Input
- Text (required) — the prompt or question
- Image 1–3 (optional) — up to three images, read together with the prompt
Parameters
Parameter | Details |
max_new_tokens | 1–8192, default 1024 (advanced). Cap on response length. Raise it for long answers or when thinking is on |
temperature | 0–2, default 0.0 (advanced). Greedy and deterministic — the model card's recommended setting for this size. Above 0 turns on sampling |
top_p | 0–1, default 0.95 (advanced). Ignored at temperature 0 |
top_k | 1–100, default 100 (advanced). Ignored at temperature 0 |
downsample_mode | 16x (default) / 4x (advanced). How much the vision encoder compresses each image. 16x is fast and fine for ordinary questions. 4x keeps 4× more visual tokens for small text and fine detail, at more time and memory per image |
enable_thinking | On / Off (default, advanced). On writes a reasoning block first — can help on multi-step visual questions; raise max_new_tokens with it |
Output
- Text — the generated response
Technical Details
Vision encoder | SigLIP2, with intra-ViT early token compression (from LLaVA-UHD v4) |
Language backbone | Qwen3.5 / Qwen3.6 family (per OpenBMB's acknowledgements) |
Size | ~1B class (Hugging Face lists ~2B including the vision tower) |
Compression | 16× (default) or 4×, selectable per request |
Upstream inputs | Single image, multi-image and video. This node exposes text and three image ports; for video use the Video Analysis (Gemini) node |
Runtime | GPU, cloud — no local setup |
Compliance & Provenance
Provider | OpenBMB |
Provider type | Open-source |
License | Apache 2.0 |
EU AI Act risk class | Limited Risk |
Art. 50 transparency | Applicable — multimodal LLM output carries the AI-generated label |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.
Model Source
- Model card
- GitHub: https://github.com/OpenBMB/MiniCPM-V
- License: Apache 2.0
How MiniCPM-V 4.7 1B compares
Model | Best for | Notes |
MiniCPM-V 4.7 1B | Fast, high-volume image questions | Smallest; up to 3 images; 16×/4× modes |
Gemma 4 E2B | Lightweight tasks | Smallest Gemma 4 |
MiniCPM-V 4.7 35B A3B | Same family, much stronger | MoE, ~3B active |
Mage-VL | Documents and charts | 4B; DocVQA 95.14 |
Gemma 4 31B | General reasoning default | MMLU Pro 85.2% |
Qwen3.8-27B | Very long inputs | 262K context |
Rough guidance: MiniCPM-V 4.7 1B or Gemma 4 E2B when speed and volume matter, the 35B A3B sibling when the answer needs more depth. For pure text extraction, an OCR model is cheaper and more accurate. Scope notes come from each node's own documentation; the models have not been benchmarked against each other inside CNAPS Studio.
All of these run inside CNAPS Studio — compare them side-by-side.
