Elorian launches Glimpse model with visual inference that re-examines the image instead of recalling it vaguely

The idea for the new model began at a game table on a holiday evening, when Andrew Dai, a former researcher at Google DeepMind, photographed a game board he was playing with friends and asked one of the leading models who was about to win, but the result was a complete inability to answer. This simple incident reveals a fundamental shortcoming that afflicts most multimodal models; they handle images with a single glance that immediately converts them into textual tokens and then perform inference in language, much like a person trying to describe a complex painting based on abstract memory after closing their eyes. From this paradox, Elorian, based in Palo Alto, launched its first models under the name Glimpse, positing a technical hypothesis that current AI does not actually see the scene despite its well-known skill in writing code and analyzing text.

The new model goes beyond reliance on textual memory by continuously returning to the original scene.During reasoning, Glimpse can crop a specific part of an image to enlarge it, retrieve particular frames from a video clip, or even execute code to measure dimensions and compare them within an environment the company calls the visual draft space. This mechanism resembles a person bringing a picture closer with their phone to verify what they saw at first glance. Dai based his approach on humans’ ability to assemble furniture using illustration manuals that contain no written explanations, wondering why computational models fail to read the shared visual symbols and cues that an adult mind understands with minimal effort.

Elorian emerged from stealth mode six months ago with a seed round of $55 million and a valuation that reached $300 million before any commercial product was launched, backed jointly by Striker Venture Partners, Menlo Ventures and Ultimeter Capital, with participation from Nvidia and Google senior scientist Jeff Dean. Although the idea of zooming and cropping during inference appeared at OpenAI in its 2025 models, the startup’s bet is to build the entire model around this philosophy at a low training cost. A small team succeeded in training Glimpse on only a few hundred graphics processors, making it much smaller than competing systems, with pricing of $0.60 per million input tokens and $3 per million output tokens, placing it in a mid-range tier that is cheaper than Sonnet and Soul and more expensive than lightweight models such as Haiku and Luna.

The company’s evaluations show the model outperforming advanced systems such as GPT-6 Astra in reading scientific charts and generating short video explanations.The model is currently available as a research preview through an API, while the chat interface remains on a waitlist, targeting in its first phase sectors that rely on massive video datasets, such as robotics labs, scientific research and visual content creators. Elorian plans to expand its capabilities next year to adopt what it describes as the best visual inference model, looking to move from a vision-and-analysis stage to an action stage that directly controls robot movement in the field through a deep understanding of their physical environment.

This shift puts CTOs and solution-engineering teams in the Gulf, Egypt and the Levant in front of different operational calculations. Projects that automate visual inspection, read engineering documents and analyse video streams in regional logistics warehouses have long faced costly inference bills imposed by large general-purpose models, alongside loss of accuracy due to the approximate textual processing of images. Providing a specialized visual model that is smaller and has reasonable consumption costs enables local teams to build precise inspection tools that rely on iterative verification instead of draining budgets on giant models that cannot read details. Your decision today is no longer limited to renting the largest available model or abandoning media processing, but moves toward integrating visual inference models that handle focused inspection tasks while leaving other systems to formulate final answers.