Mage VL
Mage-VL is a 4B-parameter multimodal model for image and video understanding, built on a from-scratch codec-native visual encoder (Mage-ViT) paired with a Qwen3-4B language backbone. It allocates video tokens by spatio-temporal importance, cutting token counts over 75% while extending usable context up to 8x and giving up to 3.5x faster inference than uniform frame sampling. A built-in cognition gate enables proactive, event-triggered streaming commentary on live video in one checkpoint.
- Type
- Multimodal
- Weights
- Open
- Downloads
- 27k
- Likes
- 405
- Pipeline
- image text to text
- License
- apache-2.0
Model hub
microsoft/Mage-VL
- Repo
- microsoft/Mage-VL
- Library
- transformers
- Access
- Public
Technical tags and languages
Papers
1
More from Microsoft
11
-
Aion 1.0 Instruct
Text
Aion 1.0 Instruct is Microsoft’s on-device small language model for local text intelligence such as summariza…
-
Aion 1.0 Plan
Multimodal
Aion 1.0 Plan is Microsoft’s 14B on-device reasoning and tool-calling model for fully local agentic workflows…
-
102k DL
biogpt
Text
text generation · License mit · 101,860 downloads · by microsoft
-
222k DL
BiomedCLIP-PubMedBERT_256-vit_base_patch16_224
Image
zero shot image classification · License mit · 222,109 downloads · by microsoft
-
8k DL
BiomedNLP-BiomedBERT-base-uncased-abstract
Text
fill mask · License mit · 799,514 downloads · by microsoft
-
382k DL
BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext
Text
fill mask · License mit · 382,055 downloads · by microsoft
-
127k DL
BiomedVLP-CXR-BERT-specialized
Text
fill mask · License mit · 127,478 downloads · by microsoft
-
234k DL
codebert-base
Text
feature extraction · 234,498 downloads · by microsoft
About Microsoft
1
Microsoft is a technology company that offers a wide range of software, cloud computing services, hardware, and artificial intelligence solutions.