Elorian, an AI startup based in Palo Alto, has made headlines by launching its first model, Glimpse, after successfully raising $55 million in seed funding at a $300 million valuation—before even bringing a product to market. The launch, which took place on Friday, marks a significant milestone for the company.
Glimpse is currently available as a research preview, featuring API access alongside a chat interface. The model is designed to address a notable limitation in current AI systems: their inability to truly “see.” Unlike traditional multimodal models that analyze an image once and assimilate it into text, Glimpse continuously references its source material. This unique capability allows it to crop sections of images, extract specific frames from videos, or even execute code to compare visual data—a feature Elorian dubs “visual scratch space.” This process is akin to zooming in on a mobile photo for closer examination.
The impetus behind this innovation comes from experiences shared by Andrew Dai, Elorian’s CEO and a former researcher at Google DeepMind. Dai recounts a puzzling incident during a game night where he asked a leading AI model to identify the winner of a board game from a photograph. The AI struggled, showcasing its difficulty in understanding the game’s unique visual language. “Board games each have their own visual language and iconography,” he explained, indicating that while humans can easily interpret these symbols, current AI models frequently flounder.
Dai posed an intriguing question: “Why can’t AI models assemble Ikea furniture from diagrams with no text, just as humans do?” This line of reasoning emphasizes the gap between human visual literacy and AI’s current capabilities.
Although the concept behind Glimpse isn’t entirely novel—OpenAI incorporated similar crop-and-zoom functionalities in its O-series models in 2025—Elorian claims to have built an entire model centered around this approach, achieving efficiency in both cost and performance. The company reports that Glimpse was trained on a modest number of GPUs and remains significantly smaller than its larger competitors in performance benchmarks.
In terms of pricing, Glimpse is positioned competitively at $0.60 per million tokens for input and $3 per million for output, making it more affordable than mid-tier models like Sonnet and Sol, yet pricier than entry-level models like Haiku and Luna.
Elorian asserts that Glimpse excels in areas such as scientific chart analysis and short-form video captioning, even outperforming GPT-6 Astra in these domains, with comparable results across various visual reasoning challenges.
The company initially targets sectors rich in video content, such as robotics, scientific research, media, and creative industries. Looking ahead, Elorian aims to evolve from visual comprehension to actionable insights—developing models that can interpret scenes and execute corresponding actions, which may include direct robot control. Next year, Dai expressed ambitions to create “the world’s best visual reasoning model,” one that outpaces all existing tests in visual analysis.
Editor’s Take
This launch from Elorian is a compelling advancement in the AI space, underscoring the growing importance of visual reasoning. By enabling AI to interact more naturally with visual data, Glimpse could enhance a variety of applications across industries, from media to robotics. As the technology progresses, it will also challenge developers to create solutions that effectively leverage these new capabilities, thus potentially reshaping user experiences and operational efficiencies.
Source: techcrunch.com
