Read Images and Text Together with a Compact 4B Vision-Language Model
Mage-VL is Microsoft's compact vision-language model, built on a visual encoder trained from scratch and a Qwen3-4B language decoder. Give it a text prompt and, optionally, an image, and it answers in text — describing, reading charts and documents, or reasoning about spatial layout. In CNAPS Studio this node runs Mage-VL's image-and-text mode. Apache 2.0.
What it does
Mage-VL takes a text prompt and, optionally, one image, and returns a text answer. It reasons over the words and the picture together, so you can ask it to describe a photo, read a chart or document, answer a question about a screenshot, or judge where things sit in a scene.
What sets it apart is its visual encoder, Mage-ViT, which Microsoft trained from scratch on about 100M unlabeled images and video frames rather than starting from a billion-scale web-pretrained encoder. Paired with the same Qwen3-4B decoder that Qwen3-VL-4B uses, it matches or beats Qwen3-VL-4B on most image benchmarks and is notably stronger on spatial reasoning.
Upstream, Mage-VL is also a codec-native video and proactive-streaming model. Those modes are not wired up in this node — it accepts text and an image only. For video understanding in Studio, use Video Analysis (Gemini).
Problem it solves
Use | What it helps with |
Describe & caption | Turn an image into a written description for alt text, catalogs, or search |
Charts & documents | Answer questions about charts, diagrams, and document images |
Spatial reasoning | Ask where objects are, how they relate, or which is closer — a relative strength for its size |
Classify by your own rules | Write the criteria in the prompt instead of training a classifier |
Light, high-volume workloads | A 4B-class model is a cheaper choice than the 27B–35B options when the task is simple |
Input/Output
Input
- Text (required) — the prompt or question
- Image (optional) — one reference image to reason over together with the prompt
The upstream checkpoint also accepts video and supports event-gated streaming. This node exposes the image-and-text mode only.
Parameters (all in Advanced Options)
Option | Range / default · what it does |
Max new tokens | 1–8192, default 2048. Cap on response length. Lower it for terse answers; raise it for long descriptions. Generation also stops at a natural end of answer regardless of this cap |
Temperature | 0–2, default 0.7. 0 is greedy and deterministic — best for reading text, extracting values, and factual Q&A. Higher values give more varied wording |
Top-p (nucleus) | 0–1, default 0.8. Keeps the smallest set of tokens whose cumulative probability reaches top-p. Lower is more focused |
Top-k | 1–100, default 20. Restricts candidates to the k most likely tokens. Lower is more conservative |
The main node view has no controls; the defaults work for most prompts.
Output
- Text — the generated response
Accuracy & Speed
Microsoft's reported image results versus Qwen3-VL-4B, which uses the same 4B language backbone — so the difference comes from the visual encoder:
Benchmark | Mage-VL-4B | Qwen3-VL-4B |
DocVQA-val (documents) | 95.14 | 94.69 |
ChartQA (charts) | 84.88 | 83.96 |
MMStar (general VQA) | 67.32 | 62.04 |
EmbSpatial (spatial) | 82.67 | 77.50 |
CrossPoint (spatial) | 80.00 | 26.90 |
TextVQA-val (text in images) | 77.28 | 80.55 |
Benchmarks are upstream figures from the official model card, not measured in CNAPS Studio. Microsoft's headline speed gains (up to 3.5× faster) apply to its codec-native video mode, which this node does not use.
Choosing a multimodal model
What you need | Use |
A compact model strong on spatial reasoning, charts, and documents | Mage-VL |
A compact model for light, high-volume workloads | Gemma 4 E2B |
Steady long single-pass reasoning, 262K context, two image inputs | Qwen3.8-27B |
Flagship quality with faster per-token inference | Qwen3.6-35B-MoE |
Video understanding — transcripts, timecodes, highlights | Video Analysis (Gemini) |
For pure text extraction from documents, an OCR model (GLM-OCR, PaddleOCR, Sarashina2.2-OCR, DeepSeek OCR) is cheaper and more accurate than a VLM.
Model Source
- Hugging Face: microsoft/Mage-VL
- GitHub: microsoft/Mage
- Project page: microsoft.github.io/Mage
- Paper: Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model
- License: Apache 2.0
Compliance & Provenance
Item | Detail |
Provider | Microsoft |
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.
Run it now
Open CNAPS Studio, find Mage-VL in the Component panel, connect a Text Input and an Image Loader, and ask a question about the image.