Elorian, a Palo Alto startup that secured $55 million in seed funding valuing it at $300 million prior to launching a product, has just introduced its first model—Glimpse.
Launching today as a research preview, Glimpse is designed to move beyond today’s standard multimodal models that simply describe images once they’re converted to tokens. Instead, it continually refers back to the original visual input as it reasons—cropping in, analyzing video frames, and running code to measure or compare visual data. This internal mechanism is called the “visual scratch space.”
Why Active Listening Isn’t Enough
Today’s leading models can translate images into text-based summaries, but often they miss the subtleties that a human eye handles naturally—symbol understanding, visual metaphors, or distinguishing nuanced detail without text prompts. Glimpse aims to fill that gap. For example, it identifies which player won a board game by parsing the visual layout and icons—something many large models struggle with.
The inspiration came from the founder’s own experiences. Andrew Dai, previously with DeepMind, noted that even top-tier AIs couldn’t tell who won a board game from a photo and couldn’t decipher common iconography without textual hints. Glimpse is built explicitly to address that limitation.
Building a Model Around Seeing Better
Though some features—like crop-and-zoom during image processing—have appeared in other AI systems, Glimpse is among the first to center its architecture around this kind of active visual reasoning. It was trained using a few hundred GPUs, and the team claims the final model is significantly smaller than many of its more resource-intensive peers.
Glimpse’s pricing model falls between lightweight systems and large-scale flagships. Input tokens cost $0.60 per million, while output tokens cost $3 per million. This positions the model as more accessible than mid-tier rivals like Sonnet and Esol, though still more costly than ultra-lightweight models such as Haiku or Luna.
In internal evaluations, Elorian reports that Glimpse beats GPT-6 Astra on scientific chart reasoning and outperforms many existing systems in generating video captions. It also achieves competitive results on standard visual reasoning benchmarks. These outcomes suggest its new method delivers both deeper understanding and reliable output on key tasks.
Target Use-Cases & Ambitions
Elorian sees early adopters across industries with large video or image datasets: robotics, media, creative firms, and scientific labs. As a research preview, Glimpse is accessible via API, with a chat-based interface available on a waitlist. Active users can already access results tables and detailed methodological write-ups.
Looking forward, Elorian’s roadmap includes moving from passive perception to action—coffee-making robots to lab assistants that not only interpret but act on scenes. The goal for the next year is ambitious: claim the top spot on every visual reasoning benchmark.
By rethinking visual understanding as a dynamic process rather than a static one, Elorian’s Glimpse may mark a turning point. It raises expectations for AI’s ability to interpret image content with context and detail. What comes next will be hardware scaling, real-world performance, and how well it can integrate into use-cases like robotics or medical imaging without losing precision or becoming cost-prohibitive.
