Elorian Breaks Stealth with ‘Glimpse,’ a New Multimodal Model Designed to Finally Give AI Real Sight

Palo Alto-based AI startup Elorian, which made significant waves in the venture capital community by securing a staggering $55 million seed round at a $300 million valuation before launching a single product, is finally pulling back the curtain. As TechCrunch has exclusively learned, the company released its first model, dubbed “Glimpse,” on Friday. The model is currently available as a research preview, offering developers both API access and a dedicated chat interface to test its capabilities.

The release marks a pivotal moment for Elorian, a company that has spent the last six months operating in a state of high-profile stealth. The core thesis behind Glimpse is as simple as it is ambitious: despite the rapid advancements in large language models (LLMs) and the impressive fluency they display in text and code, today’s AI models are fundamentally "blind." They possess the ability to describe the world, but they lack the capacity to truly see and understand it with the nuance and iterative precision of a human observer.

To understand why this distinction matters, one must look at how existing multimodal models process visual information. Most current systems operate on a linear, one-and-done approach. They ingest an image, convert the pixels into a series of tokens, and then reason about those tokens using text. It is a process akin to looking at a photograph, putting it away, and then attempting to describe it from memory. If the model misses a detail or misinterprets a subtle visual cue during that initial conversion, it has no way to correct its error. It is locked into its initial impression.

Glimpse breaks this paradigm by continuously returning to the source material. During the reasoning process, the model can dynamically crop into specific regions of an image, extract individual frames from a video, or even execute code to measure or compare elements within its field of view. Elorian refers to this mechanism as a “visual scratch space.” In practical terms, it mimics the human behavior of pinching to zoom on a smartphone screen to verify a detail—a second look that confirms or corrects a first intuition.

The impetus for this development, according to Elorian CEO and former Google DeepMind researcher Andrew Dai, came from a surprisingly mundane source: a casual board game night. Last year, while enjoying an evening with friends, Dai took a photograph of a board game that had just concluded. He then fed the image into a leading frontier model and asked it to determine who had won. The model failed. It could identify that there were pieces on a board, but it could not synthesize the complex, symbolic visual language of the game to determine the game state.

For Dai, this failure highlighted a profound gap in AI development. “Board games each have their own visual language, each have their own iconography, symbols,” Dai told TechCrunch. “To an adult, it’s quite easy to understand what’s going on with minimal text, but these models aren’t able to intuit what the symbols mean.”

He extended this observation to everyday tasks, noting that humans have an innate ability to decipher complex, non-textual instructions, such as assembling IKEA furniture by following diagrammatic cues. The fact that state-of-the-art AI models struggle with these same tasks suggests that they are not actually "seeing" the world, but rather pattern-matching against text-heavy training data. Elorian’s mission is to bridge this gap by prioritizing visual reasoning over simple text-based tokenization.

While the "crop-and-zoom" methodology is not entirely unprecedented—notably, OpenAI integrated similar capabilities into its o-series models earlier in 2025—Elorian’s competitive edge lies in its architecture. The company claims it has built the entire model around this visual-first approach from the ground up, rather than bolting it onto a pre-existing LLM. This architectural focus has allowed for significant efficiency; Elorian reports that a small, lean team trained Glimpse using only a few hundred GPUs. The result is a model that is substantially smaller than the massive frontier systems it is currently being benchmarked against, yet capable of achieving superior performance in specific domains.

The pricing strategy for Glimpse reflects this lean approach. The model is positioned at $0.60 per million tokens for input and $3 per million tokens for output. This pricing tier places it between the high-end, heavy-duty models like Sonnet or Sol and the more lightweight, budget-friendly models such as Haiku and Luna. Elorian believes this balance of cost and high-level reasoning capability will make it an attractive option for developers who need specialized visual intelligence without the overhead of massive, generalized systems.

Internal evaluations provided by Elorian show promising results. The company claims that Glimpse outperforms even the GPT-6 Astra model in tasks involving scientific chart reasoning and short-form video captioning. It also maintains competitive results across a broader suite of standard visual reasoning benchmarks, suggesting that the "visual scratch space" is more than just a theoretical advantage—it is a functional improvement that yields tangible results.

Elorian’s trajectory is backed by significant industry confidence. When it first emerged from stealth six months ago, the company announced a funding round co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital. The round also saw participation from notable figures in the AI research community, including Google chief scientist Jeff Dean, signaling strong institutional and expert support for their vision.

In the near term, Elorian is targeting industries where visual data is abundant but underutilized. Robotics firms, scientific research laboratories, media organizations, and creative studios are the primary candidates for the initial rollout of Glimpse. These sectors are currently inundated with video and image data that requires human oversight to interpret; a model that can reliably "see" and "reason" about this data could unlock massive productivity gains.

However, the long-term roadmap for the startup is far more ambitious. Dai envisions a transition from passive seeing to active doing. The goal is to develop a model that can construct a comprehensive internal representation of a scene and use that representation to select, plan, and execute actions, including the direct control of robotic hardware. By the end of next year, Dai aims for Elorian to have built what he describes as “the world’s best visual reasoning model,” one that consistently leads on every major visual intelligence benchmark.

As Glimpse moves into its research preview phase, the broader AI community will be watching to see if Elorian’s bet on "visual-first" architecture pays off. If the company can successfully prove that a smaller, more specialized model can out-reason the massive, general-purpose incumbents, it could signal a shift in how the next generation of AI is designed. The era of models that can write poetry but fail at reading a board game may be coming to a close; in its place, Elorian is attempting to usher in a future where AI possesses the kind of visual intuition that humans have long taken for granted.