Palo Alto AI startup Elorian released its first model, Glimpse, on Friday as a research preview with API access and a waitlist-only chat interface. Founded by former Google DeepMind researcher and CEO Andrew Dai, the company previously raised a $55 million seed round at a $300 million valuation. Glimpse introduces a "visual scratch space" mechanism that allows the model to continuously reference source images and videos mid-reasoning by cropping regions, pulling specific frames, or running code.
Unlike standard multimodal models that convert images into tokens once and reason about them purely through text, Glimpse mimics human visual habits like zooming in to double-check details. Elorian trained the smaller model using a few hundred GPUs, achieving lower operating costs of $0.60 per million input tokens and $3 per million output tokens. According to company evaluations, the model outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning.
Backed by investors including Nvidia, Striker Venture Partners, Menlo Ventures, and Altimeter Capital, Elorian initially targets robotics firms, scientific labs, and media companies dealing with heavy video data. The startup ultimately plans to expand from visual recognition into robotic action control and aims to build the world's best visual reasoning model next year. This development matters because it introduces a new architectural approach to visual AI that could significantly improve how artificial intelligence interprets complex visual data like charts, videos, and diagrams.