Elorian Launches Glimpse to Pioneer Native Image-Space AI Reasoning
Elorian has released Glimpse, a multimodal model that attempts to solve the limitations of current vision AI by utilizing a "visual scratch space" to re-examine image data during the reasoning process rather than relying solely on initial text tokenization.
Elorian, a Palo Alto-based startup, has released Glimpse, a multimodal model designed to address the visual limitations of existing AI architectures. According to TechCrunch, current multimodal systems often struggle because they convert visual inputs into text tokens, effectively describing a photo from memory rather than analyzing the image itself. Elorian argues that this approach fails to grasp visual nuances like specific iconography or complex diagrams. To solve this, Glimpse utilizes a mechanism the company calls a "visual scratch space," which allows the model to revisit the source image, crop into specific regions, or pull frames from video during the reasoning process.
This operational shift is intended to mimic human visual processing, such as zooming in on a photograph to verify details. CEO Andrew Dai, a former Google DeepMind researcher, noted that existing models often fail to intuit the meaning behind symbols in board games or diagrams, unlike humans who can assemble furniture from visual instructions alone. By maintaining the ability to re-examine the source material, Glimpse can run code to measure or compare elements within the visual data. The company claims this architecture is efficient, having been trained on only a few hundred GPUs, resulting in a model smaller than typical frontier systems.
Elorian’s internal evaluations suggest that Glimpse delivers significant performance advantages in specific domains. According to the company, the model outperforms GPT-6 Astra on tasks involving scientific chart reasoning and short-form video captioning. While the company reports competitive results across other visual reasoning benchmarks, these claims are based on their own testing. The model is currently available as a research preview with API access and a chat interface, priced at $0.60 per million tokens for input and $3 per million tokens for output.
The startup is positioning Glimpse for industries that manage high volumes of video and visual data. Initial target sectors include robotics firms, scientific laboratories, and creative or media companies. Elorian’s strategy is to eventually transition the model from passive observation to active manipulation. CEO Andrew Dai stated that the company intends to build a system capable of constructing a model of a scene to inform actions, including the direct control of robots. For the coming year, the company aims to establish Glimpse as the leading visual reasoning model across all standard tests.
While Glimpse introduces a specialized approach to visual reasoning, the company acknowledges that similar techniques, such as crop-and-zoom functionality, were previously implemented by OpenAI in its o-series models in 2025. Elorian’s differentiator lies in building the entire model architecture around this visual-first methodology rather than adding it as an auxiliary feature. As the startup moves forward, the primary challenge will be proving these performance claims outside of controlled internal benchmarks and demonstrating the model's utility in real-world applications where visual precision is critical for automation.