Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model
Microsoft
Microsoft researchers introduced Mage-VL, a 4B-parameter vision-language model with a visual encoder (Mage-ViT) trained from scratch on 560M images and 100M video frames. It uses codec-guided sparse token selection, borrowing I-frame/P-frame structure from video codecs, to cut visual tokens by roughly 75% versus dense frame sampling, giving up to 3.5x wall-clock inference speedup. A dual-system design pairs a lightweight streaming gate with a full reasoning decoder; it matches Qwen3-VL-4B on static VQA while beating it on video understanding and spatial reasoning.
Why it matters
Targets a real deployment bottleneck, real-time video and streaming inference cost, rather than chasing benchmark scores alone, and reports that web-scale pretraining is not essential for the visual encoder.
Importance: 2/5
Efficiency-focused multimodal research from a major lab.