Elorian, a Palo Alto-based AI startup that raised $55 million in seed funding at a $300 million valuation prior to product launch, released its first model, Glimpse, on Friday. The research preview offers API access and a chat interface, aiming to address limitations in current multimodal AI systems.

Most multimodal models process images by converting them into tokens and reasoning about them textually, akin to recalling details from memory. Glimpse differs by maintaining a dynamic visual scratch space during reasoning, allowing it to crop into specific image regions, extract frames from videos, or execute code for measurements and comparisons. This approach mirrors human behavior, such as zooming into a photo to verify observations.

Andrew Dai, Elorian's CEO and a former Google DeepMind researcher, highlighted a practical example of AI's visual limitations during a board game session. When asked to assess a photographed game state, a leading AI model failed to interpret the visual context, despite humans easily recognizing game mechanics from minimal cues. Dai noted that while humans effortlessly decode visual languages like Ikea diagrams or game iconography, AI systems lack this intuitive understanding.

Elorian trained Glimpse using a small team and a few hundred GPUs, resulting in a model smaller than comparable frontier systems. Pricing is set at $0.60 per million input tokens and $3 per million output tokens, positioning it between mid-tier models like Sonnet and lightweight options like Haiku. Benchmarks indicate Glimpse outperforms GPT-6 Astra in scientific chart reasoning and short video captioning, with results matching other visual reasoning tasks.

The company emerged from stealth six months ago, supported by investors including Striker Venture Partners, Menlo Ventures, Altimeter Capital, Nvidia, and Google chief scientist Jeff Dean. Initial applications target industries with extensive video assets, such as robotics, scientific research, media, and creative sectors. Elorian plans to advance from visual reasoning to action execution, aiming to build a model that interprets scenes and directs physical systems like robots. By next year, the company intends to release a model it describes as the best in visual reasoning across standardized tests.

Aditya Mehta, a TechCrunch reporter covering AI, reported on the launch. He is affiliated with the Tarbell Center for AI Journalism and holds a degree from UC Berkeley. Contact information is available via email or encrypted messaging.