Gemini and MM-1 researchers launch Glimpse to fix AI vision

If you snap a photo of a finished board game and ask a frontier AI who won, it will almost certainly hallucinate. The icons, tokens, and board states are clear as day to any human, but to modern multimodals, it is just a blur of text tokens.

Elorian, a Palo Alto startup founded by former Google DeepMind and Apple MM-1 researchers, wants to fix that. On October 16, the company launched its first model, Glimpse, rolling it out as a research preview with both API access and a chat interface.

Elorian had already raised a $55 million seed round at a $300 million valuation before launch, backed by Striker Venture Partners, Menlo Ventures, Altimeter Capital, Nvidia, and Google chief scientist Jeff Dean.

Pinch to zoom for neural nets

Most multimodal models today glance at an image once, translate it into tokens, and discard the pixels. Reasoning happens purely in text memory, like describing a crime scene from a quick peek.

Glimpse takes a different approach called a visual scratch space. Mid-reasoning, the model can actually query the original visual data. It can crop into a confusing corner, extract specific frames from a video clip, or even run code to measure distances and compare elements. It functions like a human pinching to zoom on a smartphone photo to verify a tiny detail.

CEO Andrew Dai said the idea crystallized when a top model failed to read his board game night results. If humans can assemble flat-pack furniture just by looking at wordless Ikea instruction manuals, Dai argues, AI models should be able to do the exact same thing.

Punching above its weight

OpenAI introduced crop-and-zoom steps to its reasoning models in 2025, but Elorian built Glimpse around this loop from scratch. The company claims a small crew trained the model on just a few hundred GPUs, keeping the model size far smaller than typical frontier systems.

Pricing sits squarely in the sweet spot at $0.60 per million input tokens and $3 per million output tokens. That makes it cheaper than mid-tier heavyweights like Claude Sonnet and GPT-5.6 Sol, though pricier than light models like Haiku.

According to internal benchmarks, Glimpse outperformed even GPT-6 Astra on scientific chart reasoning and short-form video captioning, alongside strong marks on complex counting and visual puzzles.

Why it matters

We think Glimpse represents a much smarter architectural direction for multimodals. Brute-forcing context windows to ingest massive high-resolution imagery all at once is wildly inefficient. Teaching a compact model to look back, crop, and inspect on demand lets a startup on a few hundred GPUs outperform massive frontier giants on dense visual reasoning.

For engineers dealing with schematics, scientific charts, and surveillance footage, this model looks instantly practical. Elorian's immediate targets are robotics firms, research labs, and media companies sitting on massive video archives. The company's stated endgame is physical action, moving from understanding Ikea manuals to having robots assemble the furniture.

While we still need independent benchmarks beyond Elorian's own tests, Glimpse proves that genuine visual reasoning doesn't require a ten-thousand-GPU cluster. It just requires knowing when to take a second look.