Mage-VL (4B codec-native video VLM)

The call

The most interesting genuinely-local drop of the window, and a real architecture rather than another fine-tune. Instead of sampling video frames uniformly, Mage-VL reads the video codec itself: it keeps every anchor (I) frame patch and only the motion-salient predicted (P) frame patches, using where the codec spent its bits as the signal for where the detail is. That cuts visual tokens by over 75% and yields up to a 3.5× wall-clock speedup at matched accuracy. It is codec-agnostic — traditional H.264/HEVC via motion vectors, or a neural codec via its learned rate map, with no retraining. A single checkpoint covers image understanding, offline video, and a "proactive streaming" mode where a lightweight gate stays quiet during routine footage and wakes the full model only when something worth commenting on completes.

When not to use: Text-only work — it is a vision-language model, and a 4B decoder is not a general chat or coding model. Also not a video *generator*; this is understanding only. Microsoft reports the gains against Qwen3-VL-4B with the backbone held fixed and only the encoder swapped, which is the honest comparison, but the numbers are still vendor-run — verify on your own footage before building around the streaming gate.