Elorian

AI · Computer Vision · Multimodal AI

Launch

Elorian launches Glimpse, a visual-reasoning AI backed by Nvidia and Jeff Dean

The model keeps revisiting the actual image mid-answer instead of relying on a one-time snapshot, undercutting frontier pricing by a wide margin.

• Elorian released Glimpse on October 16 as a research preview with API and chat-interface access, its first public model since emerging from stealth six months earlier.
• Elorian was founded in 2025 by CEO Andrew Dai along with Yinfei Yang and Seth Neel, ex-Google DeepMind and Apple researchers building native multimodal models for robotics, science and media.
• Glimpse's "visual scratch space" lets it crop into image regions, pull specific video frames, or run code to measure what it sees mid-reasoning, rather than converting an image to tokens just once.
• Glimpse costs $0.60 per million input tokens and $3 per million output tokens, cheaper than mid-tier rivals like Sonnet, and was trained by a small team on just a few hundred GPUs.
• The launch follows Elorian's $55 million seed round at a $300 million valuation in April 2026, with Nvidia and Google chief scientist Jeff Dean backing the company individually.
• Elorian says its own evaluations show Glimpse beating GPT-6 Astra on scientific-chart reasoning and short-form video captioning, though those benchmark claims haven't been independently verified.
• The launch tests whether a smaller, cheaply trained model built around a single architectural idea can rival frontier labs, a bet that method may matter more than raw compute scale.