Image: METAL
Summary
- Microsoft has released the image-text-to-text model Mage-VL on Hugging Face
- It reportedly supports not only image but also video understanding and streaming conversation features
- It is distributed under the Apache 2.0 license, making it immediately usable with Transformers, vLLM, SGLang, and more
Microsoft has released the vision-language model Mage-VL on Hugging Face. The model performs image-text-to-text tasks and has been confirmed to support video-understanding capabilities alongside image processing.
The model card carries tags such as multimodal, vision-language-model, streaming, and conversational, suggesting a design aimed not just at simple image question-answering but also at continuous conversational interaction and video-based input processing. The related research is reportedly detailed in arXiv paper 2607.24904.
Deployment
Mage-VL is released under the Apache-2.0 license, allowing commercial use. It can be loaded directly as a pipeline through the Hugging Face Transformers library, and server deployment via vLLM and SGLang is also supported. The model requires custom code execution (trust_remote_code), suggesting it uses a proprietary implementation rather than a standard architecture.
Aside from the fact that it was distributed through a Microsoft account with 322 likes and 21,500 followers on Hugging Face at the time of release, specific parameter counts or benchmark performance were not confirmed in the excerpted model card content.
Comments