Elorian launches Glimpse, a vision model that goes back over images and video as it reasons

發生什麼事

Elorian positions Glimpse as a model that thinks visually. Before answering, it works on the picture as part of its reasoning, for example following one feature to see how it changes, zooming in on a corner of an image, or slowing down a short stretch of video to look closely. It can also call multimodal tools on its own or split a problem among several subagents (Elorian blog).

In the comparison table Elorian published, each model wins some. Glimpse comes out ahead of Astra on four tasks (robot manipulation events, open-vocabulary counting, short-video captioning and scientific charts), ties on embodied spatial reasoning, and trails on dense video captioning and temporal reasoning (Elorian blog).