Palo Alto AI startup Elorian, which raised a $55 million seed round at a $300 million valuation before launching its product, released its first model on Friday, TechCrunch has learned exclusively.
This model, Glimpse, is starting as a research preview with API access, and the chat interface is waitlist-only, so for now it’s results tables and papers. What about the paper? Today’s AI models are still invisible, even with all the language and code skills.
Most multimodal models look at an image once, convert it into tokens, and then reason about it with text. This is similar to describing a photograph from memory. Glimpse keeps going back to the source. During inference, you can crop regions, extract specific frames from the video, or run code that measures or compares what you’re seeing. This is a mechanism Elorian calls “visual scratch space.” Imagine someone pinching a photo on their phone and zooming in to double-check what they thought they saw.
For former Google DeepMind researcher and Elorian CEO Andrew Dai, one of the clearest signs that AI still doesn’t really know comes on game night.
Last year, he photographed a board game he had just completed with some friends and asked the outlying model who had won. I couldn’t say that. Board games “have their own visual language, their own iconography and symbols,” he told TechCrunch. “It’s very easy for adults to understand what’s going on with minimal text, but these models can’t intuitively understand what the symbols mean.”
“Humans can look at an IKEA diagram and assemble IKEA furniture without text, right? Why can’t an AI model?” he said.
This approach is not completely new. OpenAI added crop and zoom during inference to its o-series models in 2025. But Elorian’s selling point is that they built an entire model around it cheaply. The company says a small team trained Glimpse on hundreds of GPUs, and its models are significantly smaller than the Frontier systems it benchmarks.
Glimpse charges $0.60 per million tokens and $3 per million tokens. It’s cheaper than mid-range models like Sonnet and Sol, but more expensive than lightweight models like Haiku and Luna.
However, in Elorian’s assessment, Glimpse outperforms GPT-6 Astra when it comes to reasoning about scientific figures and captioning short videos, and it delivers comparable results on other visual reasoning tasks.
Elorian came out of stealth six months ago with a round co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean.
The initial target will be companies that rely heavily on video, including robotics companies, scientific laboratories, media and creative companies. In the long term, Elorian wants to move from seeing to doing. This means building a model of the scene using it to select actions, such as controlling the robot directly. Dai said the company aims to build “the world’s best visual reasoning model” next year that will dominate any visual reasoning test.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
