Follow live video and react to it in real time

Mage-VL is a compact vision-language model built for video that is still happening. You can use it like a normal image and video Q&A model: give it a picture or a clip and ask what is going on. The difference shows in streaming mode. A small internal gate watches the feed, stays silent through routine content, and triggers a response only when an event completes. Microsoft's showcase is live soccer: the model produces commentary on broadcast footage, including 2026 World Cup matches it never saw in training.

The speed trick is borrowing the video codec's homework instead of studying every frame equally. Mage-VL reads a stream the way it is stored: full detail on keyframes, and only the changed regions in between. That drops roughly three quarters of the visual tokens, which the team reports as up to 3.5x faster video inference and clearly stronger scores on timing-sensitive video benchmarks than same-size rivals.

The official Space is the fastest path: upload a clip, ask a question, then switch to streaming mode and watch when the model chooses to speak. For a first test, feed it a few minutes of event-heavy video such as a match, a game session, or a dashcam clip, and check whether its interjections land on the moments you would have picked yourself.

The streaming gate was tuned on event-shaped video like sports, so on quiet footage its sense of "worth mentioning" may not match yours. The benchmark wins are Microsoft's own runs, and on text-heavy images the same-size Qwen3-VL still reads slightly better. Real-time local use needs a GPU; the browser demo is the honest first test.

• Ask questions about a long recording without scrubbing through it first.
• Generate live commentary over a match or a stream.
• Watch a feed and get a note only when something actually changes.
• Pull timestamped highlights out of hours of footage.
• Use it as a regular image Q&A model for screenshots, documents, and charts.

Try the demo

View model page

Read this issue on neuralexpedition.com →