Mate最新发布了lama3.1，405B的开源模型号称超过了chatgpt4o。 几天前用公司机器跑了一下70B，用起来还可以。 今天正好测试一下8B的能力，对自用的资源搜索站进行分类。选择最小的模型原因是要对百万级的数据进行分类，速度远比准确更重要。 Magnet SearchMagnet Search: self host torrent search, just for me. 根据种子资源title，分成几大类： Category = ["Movie","TV Show","Music","E-book","Comics", "Audiobook","Software","XXX","Unknown"