Elorian, the Palo Alto startup that caught attention by raising $55 million in a seed round at a $300 million valuation before shipping a single product, finally unveiled its first model on Friday. It's called Glimpse, and the premise is direct: instead of converting images into text tokens and reasoning from that translation, the model keeps images as live references throughout the entire reasoning process.
CEO and co-founder Andrew Dai, a former Google DeepMind researcher, told TechCrunch that the problem Glimpse aims to solve has a backstory far too mundane to be academic theory: last year, he photographed a board game after a night with friends and asked a leading frontier AI model who was winning. The model couldn't answer. "Board games each have their own visual language, each have their own iconography, symbols," Dai explained. "To an adult, it's quite easy to understand what's going on with minimal text, but these models aren't able to intuit what the symbols mean."
The analogy Dai uses is even more pointed: humans can look at an IKEA diagram with no text instructions and assemble the furniture. Current multimodal AI models fail at this type of task because they convert an image into tokens once and then reason from that static representation — like someone describing a photograph from memory. If the crucial detail falls outside the initial framing, the model hallucinates, miscounts, or gets lost in spatial relationships.
What Glimpse does differently is what the company calls a "visual scratch space." During reasoning, the model can crop into a specific region of an image, extract individual frames from a video, or execute code to measure and compare visual elements. It's not just analyzing the image once; it's returning to it, checking, confirming, refining. The entire architecture was built around this ability to revisit the source, not just encode it and move on.
The result, according to Elorian's own evaluations, is a model that outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks. Priced at $0.60 per million input tokens and $3 per million output tokens, Glimpse costs 94% less than GPT-6 Astra, which charges $10 per million input tokens. Elorian says it trained Glimpse on a few hundred GPUs with a small team, making it substantially smaller than the frontier systems it benchmarks against.
The approach isn't entirely novel. OpenAI incorporated crop-and-zoom during reasoning into its o-series models in 2025, and other multimodal systems gain access to measurement tools during inference. What differentiates Elorian is that crop-and-reference isn't a plugin bolted onto a language model — it's the core of the architecture. Every decision the model makes, from the first token to the last, can consult the original image.
This has profound implications for industries that process large volumes of visual data. Elorian is already targeting robotics firms (where vision is the primary sensing modality for navigation), scientific laboratories (which interpret microscopy, satellite imagery, and complex graphs), and media and entertainment companies (which need to analyze and generate visual content). Dai's longer-term vision goes beyond "seeing" — it's building a model that uses visual understanding to choose actions, including direct robot control. "We want to build, next year, the world's best visual reasoning model," he said.
Whether the AI community embraces Glimpse as a necessary step toward a more authentic understanding of the visual world, or whether the "visual scratch space" approach remains a specialty for cases where visual precision is critical, is still to be determined. What is undeniable is that Elorian has shown the model can be trained efficiently — a few hundred GPUs, a small team, a competitive price — against the most expensive multimodal systems on the market. If Elorian's self-reported results hold up under independent evaluation, Glimpse could be the first signal that the next leap in multimodal models won't come from more parameters, but from an architecture that is more honest about how AI should actually see.
Sources: TechCrunch, Fast Company, Explainx
✓ Independent sources cross-checked and verified before publishing
