Elorian releases Glimpse, an AI model designed for iterative visual reasoning

Elorian Launches Glimpse Model

Elorian, an AI startup based in Palo Alto, has released its initial AI model named Glimpse. The model is currently available as a research preview with API access, and a chat interface is waitlist-only. Elorian previously secured $55 million in seed funding at a $300 million valuation before product launch.

Iterative Visual Reasoning Approach

Glimpse distinguishes itself from many existing multimodal models by employing a "visual scratch space." Unlike models that convert an image to tokens once and then reason textually, Glimpse continuously refers back to the original visual source. This allows it to perform actions such as cropping specific regions, extracting frames from video, or executing code to measure and compare visual elements during its reasoning process.

Addressing Limitations in AI Vision

According to Elorian CEO Andrew Dai, current AI models struggle with true visual understanding, often failing to interpret complex visual information like board game states or instruction diagrams without text. Glimpse's iterative approach aims to overcome these limitations by enabling a deeper, more contextual analysis of visual data, similar to how humans might zoom in on a photo for more detail.

Performance and Cost

Elorian states that Glimpse was trained by a small team using a few hundred GPUs and is substantially smaller than other frontier systems. The company's evaluations indicate Glimpse outperforms models like GPT-6 Astra in scientific chart reasoning and short-form video captioning, with competitive results in other visual reasoning tasks. Pricing for Glimpse is $0.60 per million input tokens and $3 per million output tokens, positioning it between lightweight and mid-level models in terms of cost.