PALO ALTO
Elorian released its first model, Glimpse, on Friday—a research preview with both API access and a chat interface. The company raised a $55 million seed round at a $300 million valuation before launching any product.
Glimpse's core differentiation: continuous re-access to source images and video during reasoning, rather than the single-pass token conversion most multimodal systems use. The mechanism, which Elorian calls a visual scratch space, allows the model to crop into specific regions, extract individual frames, or execute code to measure and compare visual elements.
Elorian CEO Andrew Dai, a former Google DeepMind researcher, used a board game identification example to illustrate the gap: frontier AI models failed to identify a winner from a photograph due to inability to interpret visual iconography. Dai cited human proficiency assembling Ikea furniture from diagrams as a visual reasoning benchmark current AI models lack.
OpenAI introduced similar crop-and-zoom features in its o-series models in 2025. Elorian says it architected Glimpse around this core capability from inception, achieving it cost-effectively with a small team training the model using a few hundred GPUs—substantially smaller than frontier systems it benchmarked.
Pricing: $0.60 per million input tokens and $3 per million output tokens. This positions Glimpse between lightweight models and mid-tier competitors in the cost hierarchy.
According to Elorian's internal evaluations, Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results across other visual reasoning benchmarks.
Elorian emerged from stealth six months ago. Seed funding was co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean.
The company is initially targeting businesses managing large volumes of video data: robotics firms, scientific laboratories, media organizations, and creative companies. Longer term, Elorian is pursuing active scene interaction—building comprehensive models of visual scenes to inform and execute actions, potentially including direct robotic control. Dai said the company's objective for next year is to develop the world's best visual reasoning model.
