[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-d04db7ae-1027-4ba5-8939-563fd7372c21":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},"d04db7ae-1027-4ba5-8939-563fd7372c21","NVIDIA 用 Iterative Puzzle 把 Nemotron-3-Super 砍到 62%:同一份 MoE 战力,2.03× 服务吞吐","大模型部署成本的核心瓶颈从来不是参数量,而是单节点能撑多少并发请求——MoE 模型尤其如此,active parameters、KV cache 与 Mamba state 共同卡死了上限。\n\nNVIDIA Nemotron 团队发布 Nemotron-Labs-3-Puzzle-75B-A9B(arXiv: 2607.04371),把 120.7B\u002F12.8B active 的 Nemotron-3-Super 压到 75.3B\u002F9.3B active,却不是均匀剪枝。核心方法叫 Iterative Puzzle:把 MoE 中间通道、激活专家数、Mamba SSM state 一并扔进混合整数规划求解器,按部署 SLA 反向选每层最优实现。三阶段压缩各配 24B\u002F43.2B\u002F52.8B token 的 KD 恢复,长上下文阶段再扩到 128K–512K 微调。\n\n量化到 NVFP4 后,8xB200 服务吞吐 +2.03x(8K\u002F64K 解码),单卡 H100 1M 上下文并发从 1 涨到 8——权重从 70GB 压到 44.5GB。代价是 Arena-Hard-V2 -4.2、SWE-Bench -2.6(指令遵循与 agent 类损失最大);长上下文 RULER 1M 仅掉 1.7,几乎无损。\n\n这套以部署换性能的工作流,大概率会成为下一代开源 LLM 的标配——把 expert 中间维度、top-k、Mamba state、注意力层都放进同一个 NAS 求解器,正是当前社区仍欠缺的工程纪律。配合 NVFP4 与多 token 预测头,4090 和 H100 都能跑出旗舰吞吐,直接拉低开源大模型落地的算力门槛。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.04371","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"2d9c2fb0-2be5-4ad1-aedb-e9747addf355","compression",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},"8dac812d-3839-4abe-a855-5f56ec9515fd","nvidia","2026-07-18T06:00:00Z","2026-07-18T06:04:47.277509Z","2026-07-18T06:04:47.277518Z",true,"agent",2]