Editorial illustration for Elorian’s first AI model uses visual scratch space to reason with images
Elorian Unveils Glimpse: AI Model With Visual Reasoning
Elorian’s first AI model uses visual scratch space to reason with images
Elorian raised $55 million before it had a product to show for it, pricing the Palo Alto startup at $300 million on little more than a pitch. On Friday, the company finally released something: a model called Glimpse, out now as a research preview with API access. The chat interface is waitlist-only, so anyone outside the company is left with a results table and a stated thesis.
That thesis targets a specific gap in how current AI systems work. Most multimodal models look at an image exactly once, turn it into tokens, then do all their reasoning in text, the way someone might describe a photo from memory rather than keep glancing back at it. Elorian's bet is that this approach breaks down in situations that demand real visual precision, and that fixing it means letting a model return to the image itself mid-thought, not just encode it and move on.
CEO Andrew Dai, who previously worked at Google DeepMind, traces the problem to something far more mundane than a benchmark score: a photo he took of a board game last year, and a frontier model's failure to tell him who was winning.
Most multimodal models look at an image once, convert it to tokens, and then reason about it in text — like describing a photo from memory. Glimpse keeps going back to the source. Mid-reasoning, it can crop into a region, pull specific frames from a video, or run code to measure or compare what it’s looking at, a mechanism Elorian calls a “visual scratch space.”
Why this matters
Elorian raised $55 million before shipping a single model, and now we finally get to see what that bet was on: a bet that image reasoning, not language, is the next bottleneck worth solving. The pinch-to-zoom framing is a good one for developers to sit with. Most vision-language models treat an image as a static blob of tokens, encoded once and never revisited, which is exactly why they hallucinate counts, miss small text, and fumble spatial relationships. A model that can crop, pull frames, and run code mid-reasoning is at least architecturally honest about how actual visual verification works.
That said, a results table and a waitlisted chat interface is not a product, and we've seen plenty of "research preview" launches overpromise relative to what ships widely six months later. Founders building anything that depends on fine-grained visual grounding, document QA, robotics perception, video analysis, should watch whether Glimpse's API access turns into real benchmarks against GPT-4V-class models, not just Elorian's own framing of the problem.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives Cal AI's Yadegari notes influencer posts drove 1,000 new users in two days.
From Archives Meta releases Muse Spark Open, an open-weight AI model designed for Macs
Common Questions Answered
What is Elorian's visual scratch space and how does it differ from traditional multimodal models?
Visual scratch space is a mechanism that allows Elorian's Glimpse model to revisit and interact with images during reasoning, rather than converting them to tokens once and reasoning about them from memory like traditional multimodal models. During mid-reasoning, Glimpse can crop into specific regions, pull frames from videos, or run code to measure and compare visual elements, enabling more accurate image analysis and reducing hallucinations.
Why did Elorian raise $55 million before releasing a product?
Elorian raised $55 million at a $300 million valuation based primarily on its pitch and thesis about solving image reasoning as the next bottleneck in AI systems. The company's founders convinced investors that their approach to visual reasoning represented a significant opportunity worth funding before demonstrating a working product.
What specific problems does Glimpse's visual scratch space solve that current vision-language models struggle with?
Current vision-language models treat images as static blobs of tokens that are encoded once and never revisited, which causes them to hallucinate counts, miss small text, and fumble spatial relationships. Glimpse addresses these issues by allowing the model to repeatedly reference and analyze specific regions of images, improving accuracy on detailed visual tasks.
How can developers currently access and test Elorian's Glimpse model?
Glimpse is available as a research preview with API access for developers, though the chat interface is currently waitlist-only. Outside of the company, developers can review the model's performance through a results table and Elorian's stated thesis about its approach to visual reasoning.
Further Reading
- Former DeepMind researchers bet on visual AI with new Palo Alto startup - The Mercury News
- Elorian Raises $55M to Scale Visual Reasoning AI - Ventureburn
- A Former DeepMind Researcher Just Raised $55 Million Before Building Anything - Startup Fortune
- What If AI Could Actually See? - Roko's Basilisk
- Elorian AI: Ex-DeepMind Researchers Bet $55M on Visual Reasoning - Ascendants
