Most multimodal AI models look at an image once, compress it into tokens, and then reason about it in text — like describing a photograph from memory. Elorian, a Palo Alto startup led by former Google DeepMind researcher Andrew Dai, thinks that is why AI still cannot really see. Its first model, Glimpse, is designed to go back to the source image mid-reasoning.
Launching as a research preview with API access and a chat interface, Glimpse uses what Elorian calls a “visual scratch space”: while working through a problem, it can crop into a region, pull specific frames from a video, or run code to measure and compare what it is looking at — the machine equivalent of pinching to zoom on a phone photo to double-check what you thought you saw.
The approach is not entirely new — OpenAI added crop-and-zoom reasoning to its o-series models last year — but Elorian says it built the whole model around the idea, cheaply, training it on a few hundred GPUs with a small team. According to the company’s own evaluations, Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning. Pricing starts at $0.60 per million input tokens and $3 per million output tokens.
Elorian came out of stealth six months ago after a $55 million seed round at a $300 million valuation, co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with Nvidia and Google chief scientist Jeff Dean participating. The startup is targeting companies sitting on large video libraries — robotics firms, labs, media and creative companies — and aims to move from seeing to doing, including direct robot control. Meanwhile, the current generation of multimodal assistants keeps improving — see how the two leaders stack up in our ChatGPT vs Claude comparison — and image generators like Midjourney show how far visual AI has already come, alongside chat-capable models such as ChatGPT and Gemini.
Source: TechCrunch.
