Elorian’s first AI model wants to reason with images, not just about them

Elorian, a Palo Alto AI startup that raised $55 million in seed funding at a $300 million valuation before launching a product, has released its first model, Glimpse, as a research preview with API access and a waitlist-only chat interface. The model challenges the notion that current AI systems can truly see; instead of converting images to tokens once and reasoning in text, Glimpse uses a "visual scratch space" that allows it to crop into regions, pull specific video frames, or run code to measure and compare visual elements mid-reasoning. CEO Andrew Dai, a former Google DeepMind researcher, cites everyday examples like AI failing to determine who is winning a board game from a photo or assemble IKEA furniture from diagrams as evidence of this limitation. While the approach of cropping and zooming during reasoning is not entirely new, Elorian built its entire model around it cheaply, training on a few hundred GPUs with a small team. Glimpse costs $0.60 per million input tokens and $3 per million output tokens, cheaper than mid-level models but more expensive than lightweight ones. According to Elorian’s evaluations, Glimpse outperforms even GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks. Initial targets are companies with large video datasets, such as robotics firms, scientific labs, and media companies. Longer term, Elorian aims to move from seeing to doing, including directly controlling robots, and next year plans to build "the world’s best visual reasoning model."

(Source：TechCrunch)