Elorian 发布首个模型 Glimpse：可在思考中回看图像和视频
一家叫 Elorian 的公司出了首个模型 Glimpse，最特别的是边想边回头重看视频细节，自测比 GPT-6 Astra 便宜七倍还更准。
初创公司 Elorian 发布了首个多模态模型 Glimpse，特点是能在推理过程中回头重新分析图像或视频的局部：可以跳回视频中某个时间点、放大细节，或以更高帧率重放半秒画面再作答。在 Elorian 自测中，Glimpse 在科学图表阅读和短视频字幕生成两项成绩超过 GPT-6 Astra。成本方面，机器臂放置奶油芝士的规划模拟任务上，Glimpse 花费 0.035 美元，而 GPT-6 Astra 为 0.249 美元。迷宫测试演示中 Glimpse 解开 5 个不同难度的迷宫，Gemini 3.8 Flash 只解开 1 个。
Elorian has released Glimpse, its first AI model, which can go back and reanalyze part of an image or video mid-thought.
Glimpse can return to a specific time in a video while it thinks, zoom in on a detail, or replay half a second of video at a higher frame rate before answering.
> On Elorian's own tests, Glimpse scored better than GPT-6 Astra on scientific chart reading and short-form video captioning.
> For planning a simulation of a robot arm placing cream cheese in a bowl, Glimpse cost $0.035 vs. $0.249 for GPT-6 Astra.
> In the demo below, Glimpse solved 5 mazes of increasing difficulty, while Gemini 3.8 Flash solved 1/5.
