Summary
AI startup Elorian has introduced Glimpse, its first AI model designed to improve how artificial intelligence interprets and reasons about images, charts, and videos. Released as a research preview with API access and a chat interface, the model uses an approach that allows it to revisit visual information during the reasoning process.
The Palo Alto-based company believes this method could help address limitations in existing multimodal AI systems, particularly when tasks require interpreting symbols, examining visual details, or understanding information across video frames.
Elorian is initially targeting businesses in robotics, scientific research, media, and creative industries that work with large amounts of visual data.
What Is Elorian’s Glimpse AI Model?
Glimpse is a visual reasoning model developed by Elorian, an AI startup led by former Google DeepMind researcher Andrew Dai.
Unlike approaches that primarily convert an image into tokens and then reason about the resulting representation, Glimpse is designed to revisit the original visual input while working through a problem.
According to the company, the model can examine specific image regions, extract individual video frames, and run code to measure or compare visual information. Elorian calls this capability a “visual scratch space.”
The idea is similar to how a person zooms into a photograph to inspect a small detail before making a decision. Rather than relying entirely on an initial interpretation, the model can return to the visual evidence when it needs more information.
Elorian believes this approach can help AI systems handle tasks that require more than describing what appears in an image.
Why Visual Reasoning Remains a Challenge for AI
Modern AI models can describe photographs, answer questions about images, and analyze many types of visual content. However, accurately interpreting complex visual information remains challenging, particularly when the task depends on small details, unfamiliar symbols, or relationships between objects.
Andrew Dai illustrated this limitation with an example involving a board game. After photographing a game he had played with friends, he asked an advanced AI model to determine who was winning. According to Dai, the model struggled to understand the game’s symbols and visual rules.
This example highlights the difference between recognizing objects and understanding what those objects mean within a particular context.
Similar difficulties can arise when AI systems interpret scientific charts, technical diagrams, or video footage. A model may recognize individual elements but still struggle to combine them into a reliable conclusion.
Elorian’s approach is intended to address this problem by allowing the model to examine visual evidence more closely as it reasons.
How Does Glimpse’s Visual Scratch Space Work?
The visual scratch space is central to Glimpse’s design. It allows the model to perform additional visual operations during its reasoning process.
The approach includes several capabilities:
- Image cropping: The model can focus on a specific region to inspect small objects, symbols, or other details.
- Video frame selection: It can retrieve particular frames to examine relevant moments in a video.
- Code-based analysis: It can run code to measure, compare, or investigate visual information.
- Repeated visual inspection: It can revisit the source material when the initial interpretation is insufficient.
These capabilities could be useful when answering questions that require precise visual analysis rather than a general description.
For example, a scientific researcher might need to interpret values in a chart, while a robotics company could need to understand an object’s position or a sequence of movements in video footage.
The usefulness of the approach will ultimately depend on how accurately Glimpse performs these tasks across different real-world scenarios.
How Is Glimpse Different From Other AI Models?
The broader AI industry has been working on ways to improve the visual reasoning capabilities of multimodal models.
OpenAI introduced crop-and-zoom capabilities during reasoning in its o-series models in 2025, demonstrating that repeated inspection of visual information is not an entirely new concept.
Elorian’s distinction is that it says Glimpse was designed around this approach from the beginning. The company also claims it trained the model using a relatively small team and a few hundred GPUs, with a model substantially smaller than the advanced systems used in its comparisons.
This could make the approach interesting for businesses looking for visual reasoning capabilities without relying exclusively on the largest AI models.
However, model size and training resources alone do not establish overall performance. Accuracy, inference costs, speed, reliability, and the ability to handle unfamiliar tasks are also important factors.
Glimpse Performance and Pricing
Elorian says its internal evaluations show that Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning. The company also reports competitive performance on other visual reasoning tasks.
These are company-reported results, and independent testing would be needed to establish how Glimpse compares across a wider range of benchmarks and practical applications.
The model is available at the following published prices:
| Usage | Price |
|---|---|
| Input tokens | $0.60 per million tokens |
| Output tokens | $3.00 per million tokens |
According to the announcement, these prices place Glimpse below some mid-tier AI models while remaining more expensive than certain lightweight models.
For developers, the total cost will also depend on how much visual information is processed, how frequently the model revisits images or video frames, and how many output tokens a task requires.
Businesses evaluating Glimpse will need to consider both its API costs and the quality of its results for their specific workloads.
Elorian’s Funding and Background
Elorian emerged from stealth approximately six months before the Glimpse announcement, with a $55 million seed funding round at a reported $300 million valuation.
The round was co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital. NVIDIA and Google chief scientist Jeff Dean also participated, according to the original report.
The funding provided the company with resources to develop its visual reasoning technology before launching its first product publicly.
Glimpse now gives developers and businesses an opportunity to evaluate the company’s approach through its research preview, API, and chat interface.
Who Could Benefit From Glimpse?
Elorian is initially targeting organizations that handle large volumes of visual information.
Robotics companies: Robots need to interpret their surroundings and identify relevant objects and movements. Better visual reasoning could help developers build systems that understand scenes more effectively.
Scientific laboratories: Researchers often work with charts, experimental images, and visual measurements. AI tools capable of examining these materials could assist with analysis and interpretation.
Media organizations: Video analysis and captioning could help teams review footage, organize content, and identify relevant moments.
Creative businesses: Companies working with images and video could explore visual reasoning for content analysis and other production workflows.
These are potential applications rather than guaranteed outcomes. Organizations will need to evaluate the model’s accuracy and reliability before using it in important workflows.
Elorian’s Long-Term Vision for Physical AI
Elorian’s ambitions extend beyond interpreting images and videos. The company wants to develop models that can use their understanding of a scene to determine what action should happen next.
This could eventually include controlling robots directly.
Such a system would need to do more than recognize objects. It would have to understand relevant relationships, select an appropriate action, and operate within the physical constraints of the environment.
That transition from visual understanding to action is a significant technical challenge. Reliable operation would require careful evaluation, particularly in environments where incorrect actions could damage equipment or create safety risks.
Andrew Dai has said the company’s goal for 2027 is to build what he describes as the world’s best visual reasoning model, with leading results across visual reasoning benchmarks.
Conclusion
Elorian’s Glimpse model introduces a visual reasoning approach that allows AI to revisit images, examine selected video frames, and use code to investigate visual information during a task.
The company hopes this design will improve performance on problems involving scientific charts, video, and complex visual details while offering a smaller alternative to larger AI systems.
For now, Glimpse remains a research preview, and its reported benchmark advantages will need independent validation. If the model proves effective in practical settings, it could offer developers another option for building applications that depend on more detailed visual understanding.
Elorian’s longer-term goal is to move from interpreting visual scenes to using that understanding to select actions, including in robotics. Achieving that ambition will depend on progress in both visual reasoning and reliable real-world execution.