Elorian's Glimpse is a model that reasons by zooming into images
The research-preview API is live at $0.60/$3 per M; it crops, pulls video frames and runs code mid-answer. Closed weights, and the benchmarks are its own.
Copy markdown
A 'visual scratch space,' not one glance
Most multimodal models tokenize an image once and then reason in text. Glimpse instead returns to the source mid-reasoning — cropping into a region, pulling specific video frames, or running code to measure what it sees. Elorian built the whole model around re-looking, not bolted-on crop-and-zoom.
Live now via API — $0.60 in, $3 out
It's live today as a research preview on Elorian's API at $0.60/M input and $3/M output — under mid-tier models like Sonnet and Sol, over Haiku and Luna. The chat UI is waitlist-only and there are no open weights, so you can call it now but can't self-host.
Build this weekend: an agent that checks its own eyes
Point it at dense scientific charts, tiny-print documents, or long video libraries and let it zoom into the exact spot it's unsure about instead of guessing from a thumbnail. Elorian reports beating GPT-6 Astra on scientific-chart reasoning and short-form video captioning, with robotics, labs, and media archives as first targets.
The catch: a small model, scored by its maker
Every number is Elorian's own, and its comparison table still shows losses to GPT-6 Astra on broader tasks — treat Glimpse as a visual specialist and re-run your own evals before shipping. It's a small model (trained on a few hundred GPUs) from ex-DeepMind researcher Andrew Dai, backed by Nvidia and Google's Jeff Dean.