Mage-VL
Microsoft's 4B vision-language model that reads video like a codec does, with 75% fewer visual tokens and a live-stream gate.
In short. Mage-VL is a small Microsoft model that understands images, long videos and live streams. Instead of looking at every frame, it looks closely only where something changes, which makes video much cheaper to process. It is free for commercial use under Apache 2.0.
Summary
Mage-VL is a 4-billion-parameter vision-language model from Microsoft's Mage team, published on Hugging Face on 25 July 2026 under the Apache 2.0 license. It handles images, video and live streams with one set of weights, and it includes a small "cognition gate" that decides when a stream is worth commenting on.
Instead of feeding evenly sampled frames to a pretrained image encoder, it treats video the way a video codec does. It splits a stream into anchor frames and predicted frames, keeps every patch of the anchor frames, and keeps only the patches of predicted frames where motion or new detail appears. Microsoft says this cuts visual tokens by more than 75% and speeds up inference by up to 3.5 times compared with uniform frame sampling at matched accuracy. The visual encoder, Mage-ViT, was trained from scratch on about 100 million unlabeled images and video frames, and it is also released on its own (about 0.3B parameters, MIT license). The language side is Qwen3-4B-Instruct-2507, and the configuration allows 262,144 positions. The streaming gate was trained on about 3.3 million streaming samples while the rest of the model stayed frozen. The repository also bundles a neural video codec so that the model can read codec-based video.
Microsoft's tables compare it with Qwen3-VL-4B, which uses the same language backbone. Mage-VL is ahead on most video and spatial tests, for example VideoMME 64.0 against 59.7, VideoEval-Pro 45.2 against 20.7 and VSI-Bench 64.3 against 53.3, and about level on documents and general image tests (DocVQA 95.14 against 94.69). Qwen3-VL-4B is ahead on a few, such as MV-Bench (66.7 against 65.1), TextVQA (80.55 against 77.28) and CC-OCR Doc. On OVO-Bench, a test of online video understanding, Microsoft reports an overall 64.00, the best among the streaming models in its table, while the offline Qwen3-VL-4B gets 63.00. All of these are the developer's numbers. The card gives scripts for images, frame-sampled video, H.264 and HEVC video, neural-codec video and event-gated streaming, and it points to an SGLang fork for online serving.
Good for
- Video understanding and finding when something happens in a clip, at a lower token cost than frame sampling
- Live-stream monitoring or commentary, where the gate stays silent on routine content
- Spatial reasoning over video and images in a model small enough for one consumer GPU
- Research on codec-native visual encoders, since Mage-ViT is released separately
Watch out for
- It needs custom code, Transformers 5.7 or newer, and ffmpeg for codec-based video
- On text reading in images it trails Qwen3-VL-4B on TextVQA and CC-OCR Doc
- The benchmark numbers and the streaming protocol are Microsoft's own
- The gate was tested on football broadcasts, so results on other kinds of streams are unproven
- There is no LICENSE file at the top of the Hugging Face repository, only the card's Apache 2.0 statement, and the bundled codec code carries its own MIT license
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Mage-VL (4B) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Mage-VL (4B) BF16 | 4.74B | 11 GB | 5.45 GB | 3.27 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Best open-source multimodal models in 2026 (images, video and audio)
- Best small open models you can run on a laptop
Frequently asked questions
- What is Mage-VL?
- Mage-VL is a small Microsoft model that understands images, long videos and live streams. Instead of looking at every frame, it looks closely only where something changes, which makes video much cheaper to process. It is free for commercial use under Apache 2.0.
- Can I use Mage-VL commercially?
- Yes. Mage-VL is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Mage-VL open source?
- Mage-VL is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Mage-VL on my own computer?
- Yes. The smallest variant, Mage-VL (4B), needs about 3.27 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Mage-VL released?
- Mage-VL was released by Microsoft on July 25, 2026.
- How long a context window does Mage-VL have?
- Mage-VL supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Mage-VL model card (Hugging Face) huggingface.co
- Mage repository (GitHub) github.com
- Mage-VL technical report (PDF, GitHub) github.com
- Mage-ViT model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Mage-VL-Cowork Microsoft | 4.74B | 32K | Apache 2.0 (Permissive) | 3.27 GB Laptop or phone | |
| MiniCPM-V 4.7 OpenBMB | 1.55B – 35.2B 3B active | 256K | Apache 2.0 (Permissive) | 1.07 GB Laptop or phone | |
| Qwen3-VL Alibaba Qwen | 2.13B – 236B 22B active | 256K | Apache 2.0 (Permissive) | 1.47 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone |