Elorian released its first model on Friday, a startup based in Palo Alto that secured $55 million in seed funding at a $300 million valuation before it had launched anything.
The model, called Glimpse, is available as a research preview via API. Chat access remains on a waitlist. The company argues that current AI systems, despite their command of language and code, still cannot truly see images.
Most multimodal models scan an image once, convert it to text tokens, and then reason about the description from memory. Glimpse returns to the source repeatedly. It can crop into specific regions, extract frames from video, or run code to measure and compare what it is observing. Elorian refers to this as a “visual scratch space”. It functions like pinching to zoom on a phone photo to verify a detail.
Andrew Dai, the former Google DeepMind researcher and Elorian CEO, found a clear example of this failure during a game night last year. He photographed a board game he had just finished with friends and asked a leading model to identify the winner. The model could not answer.
“Board games each have their own visual language, each have their own iconography, symbols,” Dai told TechCrunch. “To an adult, it is quite easy to understand what is going on with minimal text, but these models are not able to intuit what the symbols mean.”
“Humans can look at IKEA diagrams, with no text, and assemble IKEA furniture, right? Why can’t AI models?” he asked.
The concept is not entirely new. OpenAI introduced crop-and-zoom capabilities during reasoning in its o-series models in 2025. Elorian’s claim is that it built the entire model around this mechanism at a lower cost. The company states a small team trained Glimpse on a few hundred GPUs, noting the model is substantially smaller than the frontier systems it has benchmarked against.
Pricing for Glimpse is set at $0.60 per million input tokens and $3 per million output tokens. This sits between mid-level models like Sonnet and Sol, and lightweight options like Haiku and Luna.
According to Elorian’s evaluations, Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning. Results on other visual reasoning tasks are competitive.
Elorian emerged from stealth six months ago with funding co-led by Striker Venture Partners, Menlo Ventures and Alimenter Capital. Nvidia and Google chief scientist Jeff Dean participated in the round.
Initial targets are companies holding large volumes of video data, including robotics firms, scientific labs and media and creative organisations. In the longer term, Elorian plans to move from seeing to doing. This involves building a model of a scene to choose an action, including directly controlling robots. Next year, Dai aims to build the world’s best visual reasoning model, one that leads on every visual reasoning test.
What it means
For the people making things, this changes the workflow from describing images to inspecting them. Users can now command the system to zoom in on specific details or extract frames from a video stream to verify a fact before the model forms a conclusion. This allows for checking data in scientific charts or identifying symbols in complex diagrams without relying on a single, static description.
