Elorian launches Glimpse for active visual reasoning
Elorian debuts Glimpse, an AI model using active visual reasoning to inspect images and video dynamically instead of relying on a single visual pass.
Founded by a former Google DeepMind lead, the startup introduces a system that zooms into and inspects images during processing instead of taking a single pass.
Palo Alto-based AI startup Elorian launched its first foundation model on Friday (October 16, 2026): Glimpse. Led by Andrew Dai, who previously managed data research for Gemini at Google DeepMind, the company aims to overhaul how multimodal systems process visual inputs, introducing an architecture centered on active visual reasoning. Instead of converting an image into text tokens once and relying purely on that memory, Glimpse returns to the raw pixels whenever it needs to verify details.
The model is currently available as a research preview, pairing immediate API access with a waitlist for a web chat interface. The system accepts video files up to 15 minutes long across any native frame rate.
How dynamic visual reasoning works
Most multimodal models today operate on a single forward pass. They process an image, turn everything into numerical tokens, and then answer prompts relying only on that text-like representation. In practice, that is like asking someone to answer complex questions about a building blueprint after looking at it for five seconds, without allowing them to glance back at the sheet.
Elorian built Glimpse around what it describes as a “visual scratch space.” When the model encounters ambiguous labels, tiny text, or delicate spatial layouts, it reopens the source file. The system can crop specific regions, zoom into crowded sections, run measurement scripts, or isolate individual video frames before reaching a conclusion.
Benchmarks and the Scale AI test
In figures released by Elorian, Glimpse outperformed GPT-6 Astra in internal evaluations for scientific chart analysis and short video captioning. The startup also shared scores from Humanity’s Sixth Sense (HSS), a benchmark designed with Scale AI to evaluate physical common-sense reasoning in everyday environments.
On the HSS test, Glimpse scored 46.4% accuracy. That put it behind GPT-6 Astra’s 53.6% and well short of the 93.1% human average. Because these numbers come straight from internal runs conducted by Elorian, they still lack independent third-party confirmation.
Hundreds of GPUs and market pricing
A core part of Elorian’s pitch rests on hardware efficiency. The startup claims it trained Glimpse with a lean engineering group using only “a few hundred GPUs.” That hardware footprint is a fraction of the tens of thousands of accelerators typically used by major frontier AI labs.
The company emerged from stealth six months ago after closing a $55 million seed round at a $300 million valuation. Striker Venture Partners, Menlo Ventures, and Altimeter Capital co-led the funding, alongside participation from Nvidia and Google chief scientist Jeff Dean. Technical leadership also includes Yinfei Yang, an engineering veteran from Apple and Google.
On the API side, access costs $0.60 per million input tokens and $3.00 per million output tokens, positioning Glimpse in the middle pricing tier for commercial computer vision interfaces.
What remains to be verified
Elorian has not revealed Glimpse’s parameter count or disclosed the exact dataset used in pre-training. The model remains proprietary behind an API, with no open-weights release announced. While the founders point to robotics control as their long-term research goal, today’s release is strictly confined to analyzing still images and on-demand video files.