What it is
A codec-native streaming VLM that encodes only motion-salient video patches and adds an event gate for real-time multimodal interaction. Instead of dense frame sampling then compressing, it tokenizes video in the codec domain: full I-frames plus motion/residual-selected P-frame patches, then uses a lightweight gate to trigger the decoder only on events.
Key results
- Up to 3.5x wall-clock inference speedup vs uniform frame sampling
- Over 75% fewer visual tokens while preserving spatiotemporal context
- Mage-VL-4B matches Qwen3-VL-4B on static tasks and surpasses Phi-4-reasoning-vision-15B
Shows that pushing temporal redundancy removal into the visual tokenizer can improve streaming VLM efficiency without sacrificing broad multimodal capability.