Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model

Researchers introduced Mage-VL, a streaming multimodal foundation model built for efficient real-time perception.

Its dual-system design and Mage-ViT tokenizer selectively encode motion-heavy video regions, cutting visual token use by over 75%.

Trained from scratch on large unlabeled image and video corpora, it matches or beats strong baselines on static, video, and spatial reasoning tasks.

The model is reported to deliver up to 3.5x faster inference, making multimodal AI more practical for live applications.