Elorian Unveils Pioneering AI Model Focused on Visual Reasoning, Not Just Analysis
Elorian, the Palo Alto-based AI startup that secured a $55 million seed funding round at a $300 million valuation prior to launching any products, has unveiled its first model named Glimpse.
Glimpse is debuting as a research preview, featuring API access while the chat interface has a waitlist for users. Currently, it offers a results table and a foundational thesis. The key insight? Despite their linguistic and coding capabilities, today’s AI models struggle with true comprehension of visual data.
Unlike most multimodal models that analyze an image only once and summarize it in text, Glimpse revisits the visual source repeatedly. During its processing, it can zoom in on specific areas, extract frames from videos, or execute code to measure or compare visual elements, a process referred to as a “visual scratch space” by Elorian. This is akin to someone using pinch-to-zoom on a smartphone to verify what they perceive.
Andrew Dai, the CEO of Elorian and a former researcher at Google DeepMind, noted a stark example of AI’s limitations during a board game night.
After playing a board game with friends, he took a photograph of the setup and queried an advanced AI model about the winner. The model was unable to determine who was ahead. “Board games feature unique visual languages and iconographies,” he remarked. “To a human, understanding them with little text is straightforward, but these models cannot grasp the significance of the symbols.”
“Humans can interpret IKEA assembly diagrams without text, so why are AI models unable to?” he pondered.
While this method is not entirely unprecedented—OpenAI integrated a crop-and-zoom feature into its o-series models back in 2025—Elorian’s claim is that they constructed the entire model around this core capability efficiently. The company asserts that a small team trained Glimpse on several hundred GPUs, resulting in a model that is considerably smaller than leading systems it compares to.
Glimpse is priced at $0.60 per million input tokens and $3 per million output tokens, positioning it as more affordable than mid-tier models like Sonnet and Sol but pricier than more basic options such as Haiku and Luna.
However, Elorian’s assessments indicate that Glimpse surpasses GPT-6 Astra in tasks like scientific chart reasoning and short-form video captions, showing competitive outcomes in various visual reasoning challenges.
Elorian emerged from stealth mode six months ago, following a funding round led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with involvement from Nvidia and Google’s chief scientist Jeff Dean.
Initially, the target audience comprises companies with extensive video archives, including robotics firms, research labs, and media organizations. In the long term, Elorian aims to progress from visual processing to actionable steps, aspiring to create a model that can analyze scenes and determine actions, including direct control of robots. Dai expressed the ambition to develop “the world’s best visual reasoning model,” which will excel in all visual reasoning assessments by next year.
