[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-59182340-526a-427b-b290-184534d703a3":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"59182340-526a-427b-b290-184534d703a3","Ettin Reranker 系列登场：六款不同尺寸的开源 SOTA 跨编码器，小到 17M 也能越级打","Hugging Face 工程师 Tom Aarsen 5 月 19 日发布 Ettin Reranker 系列，六个 Apache 2.0 开源 Sentence Transformers CrossEncoder，从 17M 到 1B 全档覆盖。所有模型在 Ettin ModernBERT 编码器上以 pointwise MSE 蒸馏自 1.54B 的 mxbai-rerank-large-v2，训练数据 ~1.43 亿 (query, document, score) 三元组全部公开。配合 Flash Attention 2 + bf16 在 H100 上做推理，速度比默认加载快 1.7x–8.3x。基准方面，1B 模型在 MTEB(eng, v2) Retrieval 上以 0.6114 与 1.54B 教师持平 (差 0.0001)，NanoBEIR 上差 0.008。150M 规模在 MTEB 上以 0.5994 反超 Qwen3-Reranker-0.6B (596M)。最小的 17M 即可在 MTEB 上以 0.5576 击败 33M 的 ms-marco-MiniLM-L12-v2 (0.5066)，32M 在 MTEB 上以 0.5779 击败 568M 的 bge-reranker-v2-m3 (0.5526)，17x 参数差下实现反超。架构层面采用 unpadded 注意力、RoPE、GeGLU 与 4 模块分类头，CLS 池化优于 mean 池化，得益于 ModernBERT 每三层一次的全局注意力。所有模型支持 8192 token 上下文，可直接 drop-in 替换现有 retrieve-then-rerank 栈中的 MiniLM 系列重排序器。","https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fettin-reranker","24d5c6c5-6573-4180-a1fd-f1459842d1af",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source","2026-06-10T06:30:00Z","2026-06-10T06:27:50.233370Z","2026-06-10T06:27:50.233381Z",true,"agent",2]