Elorian releases Glimpse, a visual reasoning model with iterative image interaction

Elorian, a Palo Alto AI startup founded by former Google DeepMind researcher Andrew Dai, released Glimpse, its first AI model, as a research preview with API access and a waitlist-only chat interface. The company says Glimpse is designed to reason with images by revisiting and interacting with visual data mid-reasoning, rather than converting an image to tokens once and reasoning about it in text.

The model introduces a mechanism called a 'visual scratch space,' which allows it to crop into regions of an image, pull specific frames from a video, or run code to measure or compare visual elements during the reasoning process. Elorian compares this to pinching to zoom on a phone photo to double-check details. The startup argues that most multimodal models still struggle to 'see' in this way, citing an example where a frontier model could not determine the winner of a board game from a photograph despite clear visual cues.

Elorian claims in its internal evaluations that Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning, and delivers competitive results on other visual reasoning tasks. The company attributes this to its architecture, which is built around iterative visual interaction rather than relying solely on text-based reasoning about images.

Glimpse is priced at $0.60 per million input tokens and $3 per million output tokens, positioning it between mid-tier models like Sonnet and Sol and lightweight models like Haiku and Luna. Elorian states it trained Glimpse on a few hundred GPUs with a substantially smaller parameter count than the frontier systems it benchmarked against, aiming for cost efficiency.

Elorian emerged from stealth six months prior with a $55 million seed round co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean. The startup’s initial target customers include companies with large video datasets such as robotics firms, scientific labs, media, and creative companies. Looking ahead, Elorian aims to expand from visual reasoning to 'seeing to doing,' including direct robot control, with a goal to build 'the world’s best visual reasoning model' within the next year.