Glimpse, the first model developed at Elorian for reasoning with visual material, has entered a research preview. It can revisit image regions, choose video frames and run code for measurements while developing an answer. Intended applications at the Palo Alto company include robotics, scientific graphics and media tools. Initial access includes a research API, while the model’s exact size has not been disclosed.
According to TechCrunch, the first model from Palo Alto artificial intelligence developer Elorian is Glimpse, now in research preview. While working toward an answer, it can go back to visual material. Rather than using an image only at the start, it can crop a region, choose relevant video frames and run code to make measurements.[1]
The company describes a visual working space used during reasoning. A detail inside a scientific chart or a change across a short video can be inspected again as the answer develops. Elorian aims to serve media production, creative work, robotics and scientific applications. Earlier reasoning models have also used zooming and cropping; this preview introduces Elorian’s own system built around visual operations.[1]
Initial access includes a research application programming interface, through which software can send requests to Glimpse. Andrew Dai leads Elorian after working at Google DeepMind. According to Dai, a few hundred graphics processors supplied the training compute. Its exact parameter count has not been disclosed. Comparisons involving charts and short videos come from Elorian’s own evaluations.[1]
