Elorian Launches Glimpse Model for Visual Reasoning

Palo Alto startup Elorian has released Glimpse, its first AI model designed to process visual data through an interactive visual scratch space rather than relying strictly on text descriptions. Founded by former Google DeepMind researcher Andrew Dai, the company previously secured $55 million in seed funding at a $300 million valuation from investors including Nvidia and Google chief scientist Jeff Dean. Glimpse is currently available as a research preview featuring an API and a chat interface.

Unlike traditional multimodal architectures that convert an image into text tokens once, Glimpse continuously revisits raw visual inputs while evaluating a prompt. During its reasoning phase, the model can crop into specific visual regions, isolate frames from video streams, or execute code to compare items. Elorian trained Glimpse using a small engineering team and just a few hundred GPUs, resulting in a system significantly smaller than competing frontier models.

Despite its compact footprint, Elorian claims Glimpse surpasses GPT-6 Astra on short-form video captioning and scientific chart reasoning, while remaining competitive across broader visual benchmarks. The service costs $0.60 per million input tokens and $3 per million output tokens. This places its pricing between lightweight options like Haiku or Luna and larger mid-tier systems such as Sonnet and Sol.

For developers and practitioners, Glimpse offers a specialized alternative for media-heavy workflows that usually stall on standard multimodal models. Elorian is pitching the technology to scientific labs, media companies, and robotics developers who require precise visual comprehension. Looking ahead to next year, the company plans to build models capable of translating visual understanding directly into action, including direct physical control of robotic hardware.