🔥牛逼！性能直接碾压！Hybrid Attention终于不“缩水”了：BOSS直聘南北阁开源 Depth Hybrid Attention…

This is a dev post classified by AGK Intelligence as Other (a tool drop), kept by the Dev Radar because it carries real work, not commentary.

🔥牛逼！性能直接碾压！Hybrid Attention终于不“缩水”了：BOSS直聘南北阁开源 Depth Hybrid Attention (DHA)，短上下文平均分暴涨，KV Cache几乎零增加！ 它在混合注意力架构里加入了“跨层值混合”，让每一层都能有选择地借用更早层的信息，模型表达能力明显变强，却几乎不增加计算和缓存开销。 核心优势 🔹短上下文基准全面领先：DWA平均39.68、DDA平均40.92，均超过对应的Full Attention和Hybrid基线 🔹生成任务涨幅尤其明显（GSM8K、CMATH等） 🔹长上下文检索和损失曲线也更好 🔹完整保留Hybrid Attention的低KV Cache优势，深度银行不随上下文长度增长 快速上手 1️⃣克隆仓库：git clone https://github.com/Nanbeige/Depth-Hybrid-Attention 2️⃣查看code/目录直接集成 3️⃣有问题联系：nanbeige@kanzhun.com 
GitHub：https://github.com/Nanbeige/Depth-Hybrid-Attention #Nanbeige #DHA #DepthHybridAttention #LLM #开源 #混合注意力 #AI架构 #Boss直聘

Posted by SuSu_酥酥👅 (26.1k followers) 1 h ago · 19 likes · 1.8k views · view the original post on X. Kept by the Dev Radar as Other. Tools mentioned: depth-hybrid-attention.

More dev work like this

• 老师备课找资料，不用再一页页手动翻了。这个开源工具 tchMaterial-parser，专门用于解析和下载国家中小学智慧教育平台上的教学资源。 — @bkdgiffug
• Zerion, reintroduced. — @zerion
• OpenMed 3.2 is out. — @OpenMed_AI
• 🔥厉害了！微信视觉团队直接开源2B/4B视觉大模型 WeVisOne-VL，一个模型搞定检测+抠图！真正做到物体级 + 像素级统一感知！ — @NFT_Chen
• 「拍片哥女装」整理了 1407 张 AI 服装参考图，图片里不带人物，按春夏秋冬、派对等 8 类排列，打开网页就能翻。 — @GitHub_Daily
• We brought the MOSS family. You made it feel like a community. Thank you, @LocalLLaMAsub… — @MOSS_Global
• Starting the week at Devcon, sharing insights and a grasp of our general methodology in… — @miga_labs
• Japan’s JGB repo market is estimated at about $1.6 trillion, a market where faster… — @ZenithFdn

Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.4k posts from 5k X accounts over the last 21 days, 9.1k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-03 15:57 UTC. Full method.