Tag
Microsoft introduces Mage-VL, a codec-native streaming multimodal foundation model for image and video understanding that achieves up to 3.5x inference speedup by using a sparsity pattern inspired by video codecs, cutting visual tokens by over 75%.