[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-qwen-flash-next-gdn-qsa-architecture":3,"news-related-c4027891-42ca-4517-817a-83a48550b1bb":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"c4027891-42ca-4517-817a-83a48550b1bb","Qwen3.8-Flash-Next 开源:6B 激活参数跑赢 Opus,训练成本仅 1\u002F9","阿里 Qwen 团队开源 Qwen3.8-Flash-Next:125B MoE 主干 + 51B N-gram 嵌入,每 token 仅激活 6B 参数,官方称训练成本约为 Qwen3.7-Plus 的 1\u002F9,GDN+QSA 混合注意力架构预演 Qwen4。","8 月 26 日,阿里 Qwen 团队开源了 Qwen3.8-Flash-Next 的权重。这个模型的定位很明确:它不是又一个追跑分的旗舰,而是 **Qwen4 架构的早期预览**——官方在 GitHub 仓库里直接说明,它扮演的角色和当年 Qwen3-Next 之于 Qwen3.5 完全一样:先把架构改动开源出来,让社区在 Qwen4 全家族落地前有机会审视这些设计。\n\n## 四项系统性升级:注意力、残差、嵌入、优化器\n\n这次升级不是单点修补,而是沿四个维度同时动手:\n\n- **注意力**:GDN + QSA 混合架构。Gated DeltaNet 负责高效压缩历史信息,新增的 Qwen Sparse Attention(QSA)用一个压缩轻量索引器,在微块粒度上挑选重要上下文,大幅降低长序列注意力开销。\n- **残差**:Gated Residual(GR)把残差流拓宽到 4 条分支,用动态门控管理读写,强化跨层信息流动和训练稳定性。\n- **嵌入**:N-gram Embedding 用局部上下文查表,以极少的额外计算扩容模型容量;嵌入表可以卸载到主机内存,与模型计算异步预取重叠。\n- **优化器**:采用 Muon,围绕正交化精度、Muon 与 AdamW 的分工、融合参数切分做了细化,并为新架构重新拟合了 scaling law。\n\n## 6B 激活跑赢旗舰:数字怎么读\n\n参数配置上,主模型 125B,外加 51B 的 N-gram 嵌入,每 token 只激活 6B。官方称相比 Qwen3.7-Plus,训练成本降到约 1\u002F9,但在编码和办公任务上能力更强。\n\n第三方媒体 OfficeChai 整理的 benchmark 数据更直观:SWE-bench Pro 上 62.5 分,对比 Claude Opus 4.6 Max 的 53.4;CoWorkBench 73.9 对 68.2;IFBench 指令遵循 81.3 对 62.5,差距最大的就是这一项。当然也有输的地方:Humanity's Last Exam 上 Claude 40.0 领先 Qwen 的 35.9,前沿推理仍是闭源强项。上下文方面原生 256K,通过 YaRN 可扩到 1M,官方称 1M token 处 prefill 提速至 7.6 倍、decode 至 4.9 倍。\n\n## 所以呢\n\n一个每 token 只激活 6B 参数的模型,在阿里自己选定的 benchmark 集合上大面积追平甚至反超旗舰级闭源模型——不管你信不信这些跑分,「用 1\u002F9 训练成本逼近旗舰」这个趋势本身值得每个做推理成本规划的人认真对待。权重已在 Hugging Face 和 ModelScope 上线,SGLang、vLLM、llama.cpp、MLX 均已支持,门槛建议自己跑一遍验证。([GitHub 官方仓库](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen3.8-Flash-Next\u002F),[OfficeChai 分析](https:\u002F\u002Fofficechai.com\u002Fai\u002Fqwen-3-8-flash-next-benchmarks\u002F))","https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen3.8-Flash-Next\u002F","c36a21ac-2a77-421b-9519-1e150695732a",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"e7852287-0bea-4610-b4b9-72445589ffe6","en","Qwen3.8-Flash-Next Open-Sourced: 6B Active Params Beat Opus at 1\u002F9 Training Cost","Alibaba's Qwen team open-sources Qwen3.8-Flash-Next: a 125B MoE backbone plus 51B N-gram embeddings with only 6B parameters active per token. Training cost is claimed at roughly 1\u002F9 of Qwen3.7-Plus, previewing the GDN+QSA architecture for Qwen4.","On August 26, Alibaba's Qwen team released the weights of Qwen3.8-Flash-Next. The positioning is explicit: this is not another benchmark-chasing flagship, but an **early preview of the Qwen4 architecture**. The official GitHub repository states plainly that it plays the same role Qwen3-Next once played for Qwen3.5 — open-source the architectural changes first, so the community can examine them before the full Qwen4 family is built on top.\n\n## Four Systematic Upgrades: Attention, Residual, Embedding, Optimizer\n\nThis release is not a single-point patch. The model is upgraded along four axes simultaneously:\n\n- **Attention**: a GDN + QSA hybrid. Gated DeltaNet compresses history efficiently, while the new Qwen Sparse Attention (QSA) uses a compressed lightweight indexer to select important context at micro-block granularity, substantially cutting long-sequence attention cost.\n- **Residual**: Gated Residual (GR) widens the residual stream into 4 branches and controls reads and writes with a dynamic gate, strengthening cross-layer information flow and training stability.\n- **Embedding**: N-gram Embedding looks up a table using local context, scaling model capacity with very little extra compute; the embedding table can be offloaded to host memory and overlapped with model computation through asynchronous prefetching.\n- **Optimizer**: the Muon optimizer is adopted, refined around orthogonalization accuracy, the division of labour between Muon and AdamW, and the splitting of fused parameters, with the scaling law refitted for the new architecture.\n\n## 6B Active Params vs Flagships: How to Read the Numbers\n\nOn configuration: a 125B-parameter main model, plus 51B N-gram embeddings, with only 6B parameters active per token. Officially, training cost is about 1\u002F9 of Qwen3.7-Plus, yet the model delivers superior capability in coding and office tasks.\n\nBenchmark data compiled by third-party outlet OfficeChai is more concrete: 62.5 on SWE-bench Pro versus 53.4 for Claude Opus 4.6 Max; 73.9 on CoWorkBench versus 68.2; and the widest gap on IFBench instruction following, 81.3 versus 62.5. There are losses too: Claude leads Humanity's Last Exam 40.0 to 35.9 — frontier reasoning remains a closed-model stronghold. Context is native 256K, extendable to 1M via YaRN, with reported prefill speedups up to 7.6x and decode up to 4.9x at the 1M-token mark.\n\n## So What\n\nA model activating only 6B parameters per token broadly matches — and on a curated benchmark set Alibaba chose, beats — a frontier-class closed model. Whether or not you trust the charts, the trend of \"approaching flagship at 1\u002F9 the training cost\" deserves serious attention from anyone planning inference budgets. Weights are live on Hugging Face and ModelScope, with SGLang, vLLM, llama.cpp, and MLX support already in place — the barrier to verifying it yourself is low. ([Official GitHub repo](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen3.8-Flash-Next\u002F), [OfficeChai analysis](https:\u002F\u002Fofficechai.com\u002Fai\u002Fqwen-3-8-flash-next-benchmarks\u002F))","qwen-flash-next-gdn-qsa-architecture","2026-08-27T15:10:00Z","2026-08-27T15:10:55.826409Z","2026-08-27T15:10:55.826419Z",true,"agent",46,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"637f84e0-e6dc-490a-bba1-879f6527bdd5","Qwen3.8-Max 2.4T 开源:Gated DeltaNet 把长上下文成本砍到 1\u002F8","qwen3-8-max-2-4t-open-weights-gated-deltanet","2026-08-30T03:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"61de017b-bdd6-44b3-9f45-d4fb233bd24d","PhoneLLM 开源:30B MoE 电话客服模型,自称比 GPT-5.6 Terra 便宜 94%","phonellm-alpha-1-voice-agent-open-model","2026-08-29T21:10:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"33f3b08b-c8a2-43ec-81cf-85e2b918f913","腾讯开源 Hy4 preview:770B MoE、1M 上下文,模型首次参与自身训练","tencent-hy4-preview-770b-moe","2026-08-29T15:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"3d36921f-3b84-4663-97a0-fee7d4eff795","汤森路透开源 Thomson-1.0-Small:持续学习改造 Qwen,3B 激活的 35B MoE","thomson-1-0-small-continual-learning","2026-08-28T19:10:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"804ab59a-a8d6-4b61-bf74-8f6f2bdae83c","智谱把 Flash 做成一件正经事:一次说清 GLM-5.3-Flash 的架构和 benchmark 真相","glm-5-3-flash-hybrid-attention-architecture","2026-08-27T08:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"7958a2f1-028c-4b4e-b134-0d5de9afc1c1","Motif 3 收官:韩国 314B MoE 改用 MIT 许可,从零起步架构首次面向商用","motif-3-mit-license-sovereign-ai","2026-08-24T00:00:00+00:00"]