CEO Andrew Dai, a former Google DeepMind researcher, developed the system to address persistent failures in AI visual intuition, such as interpreting board game states or silent instructional diagrams. Unlike conventional models that convert images to text before processing, Glimpse performs iterative inspections—much like a human pinching to zoom on a photo—to verify details in real-time. The startup trained the model on a modest cluster of a few hundred GPUs, aiming for efficiency over the massive scale typical of frontier systems.
In section Startups & Technology
Elorian Launches Glimpse to Solve AI Visual Reasoning
Palo Alto startup Elorian has unveiled Glimpse, a multimodal AI model designed to move beyond static image analysis by actively querying visual data. Rather than relying on a single-pass tokenization process, the system employs a 'visual scratch space' that allows it to zoom and re-examine specific regions during reasoning.
While OpenAI integrated similar crop-and-zoom capabilities into its o-series in 2025, Elorian claims its architecture is built entirely around this dynamic approach. The company reports that Glimpse outperforms models like GPT-6 Astra on scientific chart analysis and video captioning tasks. Priced at $0.60 per million input tokens, the model targets industries reliant on high-volume visual data, including robotics, scientific research, and media production. Backed by investors including Menlo Ventures and Nvidia’s Jeff Dean, Elorian plans to evolve the technology toward active scene manipulation and direct robot control by next year.
Comments (0)
No comments yet. Be the first!