Elorian’s first AI model wants to reason with images, not just about them
techcrunch ·
Elorian, a Palo Alto-based AI startup, has released its first model, Glimpse, which is designed to reason with images rather than just describing them. Unlike traditional multimodal models that convert visuals into tokens for text-based analysis, Glimpse utilizes a 'visual scratch space' to actively zoom, crop, and interact with visual data during the reasoning process. This approach addresses limitations in current AI capabilities, such as the inability to intuitively understand board game symbols or IKEA diagrams without textual cues. The model was trained by a small team on a few hundred GPUs, resulting in a system that is substantially smaller and cheaper than frontier competitors like GPT-6 Astra. Despite being more expensive than lightweight models, Glimpse reportedly outperforms larger systems in scientific chart reasoning and video captioning tasks. Elorian initially targets industries with heavy video usage, including robotics and scientific labs, while planning to evolve toward action-oriented visual control in the future. The company launched after raising $55 million in seed funding at a $300 million valuation, backed by investors including Nvidia and Google chief scientist Jeff Dean.
techcrunch