Elorian’s Glimpse AI keeps looking at the evidence

Elorian has released Glimpse, its first AI model, built around a simple premise: a system should not have to answer visual questions from its first reading of an image or video. Many multimodal models create an initial internal summary of visual input and do most later reasoning from that; Glimpse is designed to revisit the original material, zoom into a region, pull particular video frames, or use code to measure and compare what it sees. The distinction matters when a small label, an object count, or a change across frames determines the answer—situations where a broad description can miss the point. It is a research preview for developers, so the real test is whether this repeated checking proves more reliable outside the company’s demonstrations.