AI startup Elorian has released Glimpse, a model designed to inspect images and video repeatedly as it works through a question. Rather than relying only on an initial image conversion and a text description, it can return to visual material, focus on a region or frame, and use code to measure or compare details.
A model that can look again
Many multimodal systems process an image once, turn it into tokens, and then reason in text. Elorian says Glimpse can revisit the source during reasoning. It may crop into part of an image, select particular frames from a video, or run code to examine what it is seeing.
The company calls this capability a “visual scratch space.” The idea is straightforward: a model should be able to check visual evidence while answering, much as a person might zoom in on a phone photo to verify a detail. This approach is intended to help with tasks where recognizing symbols, comparing objects or reading a chart matters.
Glimpse is being released as a research preview with API access. Its chat interface is waitlist-only, so access to the conversational product is limited for now. Elorian’s early presentation of the model emphasizes evaluation results and its underlying thesis about how AI systems handle visual information.
Why visual reasoning matters
Elorian CEO Andrew Dai, a former Google DeepMind researcher, described a board-game question as one example of the limits he has seen. After photographing a game board he had just played with friends, he asked a leading AI model to identify who was winning. It could not answer.
Dai said board games rely on their own visual language, including iconography and symbols. A person can often infer what is happening with little or no written explanation, while a model may not understand what those visual elements mean. He also compared the challenge to IKEA diagrams, which people can use to assemble furniture without relying on text instructions.
These examples point to a broader distinction: describing an image is different from reasoning through it. A system may produce a plausible account of a picture yet struggle to interpret relationships or infer what a symbol represents. Elorian’s design gives Glimpse a way to inspect evidence again during the reasoning process.
Performance, size and price
Elorian says a small team trained Glimpse using a few hundred GPUs, and that the model is substantially smaller than the frontier systems used in its comparisons. The company’s claim is that visual reasoning capabilities can be built into a model without the scale of those larger systems.
According to Elorian’s evaluations, Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning. The startup reports competitive results on other visual reasoning tasks as well. These are company evaluation claims; the source does not provide additional independent results.
The listed price is $0.60 per million input tokens and $3 per million output tokens. That is cheaper than mid-level models such as Sonnet and Sol, while more expensive than lightweight models such as Haiku and Luna. The pricing and the model’s smaller reported size may be relevant to organizations assessing how to process large volumes of visual material.
From video analysis to action
Elorian’s initial customers are expected to include businesses with extensive video collections, such as robotics firms, scientific labs, and media and creative companies. For these organizations, the ability to return to a frame or inspect a specific region could support work that depends on visual details across many recordings.
The longer-term ambition goes beyond analysis. Dai said Elorian wants to build a model that constructs an understanding of a scene and uses it to choose an action, including direct control of robots. The company’s stated goal for next year is to build “the world’s best visual reasoning model,” leading on every visual reasoning test.
Elorian emerged from stealth six months ago with a funding round co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean. Glimpse is the company’s first model, and its research preview puts its central bet into practice: visual AI should be able to look back at what it sees while deciding what it means.