The research preview can crop images, inspect video frames and use code while working through an answer.
TechCrunch reports that Elorian, a Palo Alto startup that raised a $55 million seed round at a $300 million valuation before launching a product, released its first model, Glimpse, as a research preview with API access and a chat interface. The company’s thesis is that most multimodal models look at an image once, turn it into tokens and then reason in text. Glimpse can instead crop into a region, pull specific frames from a video, or run code to measure or compare what it sees, which Elorian calls a “visual scratch space.”
The page notes the approach is not entirely new: OpenAI added crop-and-zoom during reasoning to its o-series models in 2025. Elorian says a small team trained Glimpse on a few hundred GPUs and that it is substantially smaller than the frontier systems it is benchmarked against. These are the company’s claims;
In plain English
A “multimodal” model handles more than text, such as images and video. A “token” is a chunk of text or image data the model processes. A “research preview” is an early release that may change.
Why it matters
It targets a known weakness, models misreading pictures and diagrams, by letting the model look again, which matters for any task that depends on reading charts, screenshots or product photos.
The useful distinction is whether an assistant can revisit the visual evidence when it encounters uncertainty. A chart may contain a small axis label that changes the interpretation of every number. A video may require checking a particular frame rather than describing the overall scene. These are examples of where targeted inspection could help, not verified demonstrations of this model.
For a potential user, a good evaluation would start with images whose answers are already known. Include small text, crowded diagrams and examples where the correct response is that the image does not provide enough information. Ask for the specific region or frame supporting the answer. That makes it easier to separate a convincing explanation from an accurate observation.
Extra inspection also has a cost. Cropping, running code and revisiting frames can add processing time and tool calls. The meaningful comparison is the cost of a correct, reviewable answer, including human checking, rather than the model's response speed alone.
A research preview is an opportunity to test the approach. It does not establish that the system is reliable for every visual task. Independent evaluations, clear failure examples and reproducible comparisons would make the launch claims much easier to assess.