Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model
Senqiao Yang, Kaichen Zhang, Zhaoyang Jia, Jinghao Guo et al. · arXiv · 2026
Abstract
Vision-language models handle hard offline reasoning well but are slow and clumsy at simple live perception. This model borrows from video codecs instead of sampling frames at a fixed rate: it encodes only the motion-rich regions between anchor frames, cutting visual tokens by more than seventy-five percent. A lightweight gate decides when something deserves attention and a heavier decoder handles it — a split the authors liken to fast and slow thinking.
Why it matters
Interesting for what it borrows. Video compression solved 'which pixels carry information' decades ago, and applying that to tokenisation is a more elegant fix than simply sampling fewer frames.
https://arxiv.org/abs/2607.24904