A vision model that goes back to the evidence: Glimpse zooms into details, replays video intervals, and tracks one feature across frames

A visual reasoning model can now be called through AI Gateway with video as well as still images, extending the task from reading a scene to following object motion, judging spatial relationships, and tying an action to its physical consequences. Such calls previously went through each provider's own interface, while AI Gateway offers a unified API for calling models, tracking usage and cost, and configuring retries, failover, and performance optimizations. The release gives examples of checking a train's motion to determine whether it is slowing down, keeping objects distinct through camera movement, and labeling completed actions in robot footage.

The model can return to the original visual evidence while it works, rather than relying only on a single pass over the input. That return takes three concrete forms: zooming into a detail, revisiting a short interval of video, or tracking the same feature across frames. The release lists scientific chart analysis, video captioning, visual counting, and robotics data annotation as uses this supports; calls use elorian/glimpse as the model name, send video as a file part, and can set video_fps to control frame sampling.

The release addresses developers calling the model through AI Gateway, covering image and video input, the three ways of returning to evidence, and uses such as scientific chart analysis, video captioning, visual counting, and robotics data annotation. It also gives the call format: the model name is elorian/glimpse, video is sent as a file part, and the frame rate can be set.

The described capabilities appear as example scenarios without accuracy figures or comparisons to other models; the release gives no quantitative account of how the return operations perform on long video or complex charts.