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

Palo Alto-based AI startup Elorian has unveiled its first model, Glimpse, designed to reason with images rather than just describing them. The model is currently available as a research preview, accessible via API and a limited chat interface. Glimpse aims to address a limitation in today's AI models, which, despite their proficiency in language and code, struggle to "see" as humans do. Most multimodal models process images once, transforming them into tokens and reasoning about them in text form.

In contrast, Glimpse maintains a continuous loop, allowing it to zoom in on specific regions, extract frames from videos, or execute code to measure and compare visual elements. This unique mechanism, called a "visual scratch space," enables the model to dynamically adapt and refine its understanding of an image. Andrew Dai, Elorian's CEO and a former Google DeepMind researcher, likens this capability to the human ability to visually inspect details, such as during a game night when he asked a frontier model to identify the winner of a board game from a photo.

The model failed to recognize the symbols and winning condition, highlighting a gap in AI's visual comprehension.

Although not entirely novel, Elorian's approach emphasizes building the entire model around the visual reasoning aspect, which they claim is more cost-effective. Glimpse was trained on a small team's GPUs and is notably smaller than leading frontier systems. Pricing for Glimpse is set at $0.60 per million input tokens and $3 per million output tokens, positioning it between lightweight models like Haiku and Luna, yet more expensive than mid-level models like Sonnet and Sol.

Despite its higher cost, Glimpse demonstrates superior performance in scientific chart reasoning and short-form video captioning compared to models like GPT-6 Astra. Elorian, which went out of stealth six months ago, secured a $55 million seed round co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with backing from Nvidia and Google's chief scientist, Jeff Dean.

The company aims its initial applications at industries with abundant video data, such as robotics, scientific research, and media. In the long run, Elorian envisions Glimpse as the foundation for a "visual reasoning model of the world," capable of understanding and interacting with scenes to perform actions, potentially controlling robots directly. Elorian's long-term goal is to lead in every visual reasoning test by the end of next year.