Elorian’s first AI model wants to reason with images, not just about them

Elorian, the Palo Alto AI startup that raised a $55 cardinal effect circular astatine a $300 cardinal valuation earlier it launched a product, released its archetypal exemplary connected Friday, TechCrunch has exclusively learned.

The model, Glimpse, is launching arsenic a probe preview with API entree and a chat interface is waitlist-only, truthful for present it’s a results array and a thesis. The thesis? Today’s AI models, for each their accomplishment with connection and code, inactive can’t truly see. 

Most multimodal models look astatine an representation once, person it to tokens, and past crushed astir it successful substance — similar describing a photograph from memory. Glimpse keeps going backmost to the source. Mid-reasoning, it tin harvest into a region, propulsion circumstantial frames from a video, oregon tally codification to measurement oregon comparison what it’s looking at, a mechanics Elorian calls a “visual scratch space.” Think of idiosyncratic pinching to zoom connected a telephone photograph to double-check what they thought they saw.

For erstwhile Google DeepMind researcher and Elorian CEO Andrew Dai, 1 of the clearest signs that AI inactive couldn’t truly spot came connected crippled night. 

Last twelvemonth helium photographed a committee crippled he’d conscionable finished with friends and asked a frontier exemplary who was winning. It couldn’t say. Board games “each person their ain ocular language, each person their ain iconography, symbols,” helium told TechCrunch. “To an adult, it’s rather casual to recognize what’s going connected with minimal text, but these models aren’t capable to intuit what the symbols mean.”

“Humans tin look astatine IKEA diagrams, with nary text, and assemble IKEA furniture, right? Why can’t AI models?” helium said. 

The attack isn’t wholly caller — OpenAI added crop-and-zoom during reasoning to its o-series models successful 2025 — but Elorian’s transportation is that it built the full exemplary astir it, cheaply. The institution says a tiny squad trained Glimpse connected a fewer 100 GPUs and that the exemplary is substantially smaller than the frontier systems it’s benchmarked against. 

Glimpse costs $0.60 per cardinal tokens successful and $3 per cardinal tokens retired — cheaper than mid-level models similar Sonnet and Sol but much costly than lightweight models similar Haiku and Luna.

But according to Elorian’s evaluations, Glimpse outperforms adjacent GPT-6 Astra connected technological illustration reasoning and short-form video captioning, with competitory results connected different ocular reasoning tasks. 

Elorian came retired of stealth six months agone with a circular co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with information from Nvidia and Google main idiosyncratic Jeff Dean.

The archetypal targets are companies sitting connected tons of video: robotics firms, technological labs, media and originative companies. Longer term, Elorian wants to determination from seeing to doing. That means gathering a exemplary of a country utilizing that to take an action, including straight controlling robots. Next year, Dai said, the institution aims to physique “the world’s champion ocular reasoning model,” 1 that leads connected each ocular reasoning test.