[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-moe":3},{"tag":4,"articles":11},{"id":5,"name":6,"slug":7,"article_count":8,"description_zh":9,"description_en":10},"dc48d588-7063-4e5e-9689-757acf80390d","MoE","moe",9,"MoE（Mixture of Experts）是一種讓模型按需啟用部分專家的架構，常用來在總參數、推理成本與效果之間取得平衡。從開源寫碼模型到長上下文 agent 系統，MoE 正成為大模型工程化的重要路線。","MoE, or Mixture of Experts, is an architecture that activates only a subset of experts per token or task, balancing total parameter count, inference cost, and quality. It shows up in open coding models, long-context agents, and other systems built for efficient scaling.",[12],{"id":13,"slug":14,"title":15,"summary":16,"category":17,"image_url":18,"cover_image":18,"language":19,"created_at":20},"1e4ba03d-b371-427a-8d9e-d694f09827b1","unipool-shared-expert-pool-moe-en","UniPool shares MoE experts across layers","UniPool replaces per-layer MoE experts with one shared pool, cutting redundancy and improving validation loss in five LLaMA-scale models.","research","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778221264459-eh59.png","en","2026-05-08T06:20:40.975202+00:00"]