Glimpse

Glimpse is Elorian's first model, a multimodal visual-thinking system released as a research preview. It reasons directly in visual space by revisiting and refining image/video representations during inference (e.g., cropping regions, pulling frames, running code for measurements) rather than converting once to tokens. Trained end-to-end on an open-weights base with multimodal tools and subagents, it targets grounded tasks in robotics, scientific charts, and video understanding.