New AI Model Glimpse Looks Back at Pictures While It Thinks

Elorian, an AI startup in Palo Alto, California, has released its first model, called Glimpse. The company raised a $55 million seed round (early funding) at a $300 million valuation before it had any product to sell. Glimpse is launching as a research preview, which means it is an early version for people to try, with API access for developers and a chat window.

The idea behind it is that today's AI models are very good with language and computer code but still cannot truly see. Most of them look at an image one time, turn it into small pieces of data called tokens, and then think about it using words. Elorian compares that with describing a photo from memory.

Glimpse works differently. While it is reasoning, it can go back to the picture. It can crop a section, pull certain frames out of a video, or run code to measure or compare things. Elorian calls this a "visual scratch space." It is like zooming in on a phone photo to double-check what you thought you saw.

Andrew Dai, Elorian's CEO and a former Google DeepMind researcher, says he saw the problem on game night. He took a photo of a board game he had just finished with friends and asked an advanced AI model who was winning. It could not tell him. Board games use their own symbols and pictures, and Dai said an adult can understand them easily, but the models could not. He also pointed out that people can build Ikea furniture from diagrams with no words, and asked why AI cannot do the same.

The idea is not completely new. OpenAI added a similar crop-and-zoom step to some of its models in 2025. Elorian says the difference is that it built its whole model around the idea. The company says its small team trained Glimpse using only a few hundred GPUs (powerful computer chips), and that the model is smaller than the leading systems it was tested against.

According to Elorian's own tests, Glimpse does better than a leading rival model at reading scientific charts and at writing short descriptions of videos. Its first customers are expected to be companies with lots of video, such as robotics firms, science labs, media companies and creative studios.