Elorian: glimpse

Elorian: Glimpse is Elorian's first model, released on October 16, 2026 as a research preview with API access and a waitlist-only chat interface. It accepts text, image, and video input and produces text output, backed by a 262K-token context. Its defining trait is a visual scratch space: rather than relying on a single lossy first impression, it revisits the actual images and videos mid-reasoning, zooming into image corners to inspect details, slowing video to a higher frame rate for review, and tracking a specific object across multiple frames. The model is built for visual workloads that demand close attention to detail. It suits synthetic data labeling and high-level planning in robotics, video review on construction and manufacturing sites, dense video captioning for media archives, accurate object counting on construction sites, timestamped video review, and comparing multiple images. It supports tool calling and structured outputs, and Elorian positions it for applications that perform frequent, large-volume visual analysis across both physical-world and media use cases. Elorian says it fills a gap where frontier general models are either insufficiently accurate or too expensive for visual workloads, and notes that Glimpse reports when it cannot identify an item. Independent hands-on testing tempers the picture. Glimpse does not outperform GPT-6 Astra and Claude Opus 5.5 on every task; GPT-6 Astra remains stronger on fine visual puzzle tasks and dense video captioning, while Claude Opus 5.5 is slightly better at reading scientific charts. Glimpse can also contradict its own earlier statements within long multi-part answers, so it is not fully self-consistent. Elorian's own comparisons claim Glimpse surpasses GPT-6 Astra on scientific chart reasoning, open-vocabulary object counting, short video captioning, and manipulation event understanding, and matches it on embodied reasoning, though no scores are given.