[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-qwen3-8-max-2-4t-open-weights-gated-deltanet":3,"news-related-637f84e0-e6dc-490a-bba1-879f6527bdd5":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},"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","阿里通义千问 8 月 3 日正式开源 Qwen3.8-Max 权重,2.4T 总参数 \u002F 95B 激活;同源 Flash-Next 把 Gated DeltaNet 与 Qwen Sparse Attention 嫁接,长上下文算力与显存开销显著下降。","阿里通义千问 8 月 3 日把自家最贵的 Qwen3.8-Max 端了出来,这次不仅照常发 API,还第一次承诺把 Max 级权重开源:2.4T 总参数、95B 激活,模型体量直接跨过\"开源最大 LLM\"那道坎,权重\"下周\"上 Hugging Face 与 ModelScope。\n\n## 不只是又一个大模型:架构换装\n\n相比 Qwen3.5 时期那一版 Gated DeltaNet + 稀疏 MoE 混合架构,Qwen3.8 在底层做了四件事。Flash-Next 是这一代架构的预览版:125B 主干 + 51B N-gram Embedding,每 token 只激活 6B。改动落到四个层面。\n\n- **注意力**:Gated DeltaNet(GDN)继续负责压缩历史,新加的 **Qwen Sparse Attention(QSA)** 用一个压缩过的轻量索引器在微块粒度上挑重点,把长序列注意力成本压下来。\n- **残差流**:**Gated Residual(GR)** 把残差流拓宽成 4 路,再用动态门控控制读写,跨层信息流更稳、训练也更稳。\n- **Embedding**:**N-gram Embedding** 用局部上下文查表扩展容量,embedding 表可卸载到主存、用异步预取与计算重叠。\n- **优化器**:全栈换上 **Muon 优化器**,围绕正交化精度、Muon 与 AdamW 的分工、融合参数的切分做了重新调校,scaling law 也按新架构重测。\n\n官方口径里,这套新架构的\"下一代\"位置很明确:Qwen3.8-Flash-Next 扮演的角色,和当年 Qwen3-Next 之于 Qwen3.5 一样——先放架构变化让社区检验,再把完整 Qwen4 叠上来。\n\n## 跑分:对 GPT-5.6 Sol、Fable 5、Opus 4.8 的硬碰硬\n\n官方给出的对照表覆盖了 38 项基准。挑几个硬骨头看:Terminal Bench 2.1 上 Qwen3.8-Max 拿 86.6,接近 GPT-5.6 Sol 的 88.8,压过 Opus 4.8 与 Fable 5 的 84.6;PaperBench 拿到 93.0,跑赢 GPT-5.6 Sol(90.5)、Fable 5(88.8)、Opus 4.8(80.3);AndroidBench 75.1,与 Fable 5 的 84.5 仍有差距,但超过 GPT-5.6 Sol 的 74.0。\n\n视觉侧更亮眼:BabyVision 82.0 \u002F 91.3(无 \u002F 有 Code Interpreter),把 Opus 4.8 的 28.4 \u002F 81.2 远远甩开;HLE-VL(w\u002F Tools)拿到 52.2,超过 GPT-5.6 Sol 的 51.2;ZeroBench Pass@5 拿到 24.0 \u002F 49.0,也是全场最高。多模态文档与视频理解里,OmniDocBench 1.5、视频 MME(w\u002F Sub.)、MLVU(M-Avg)、LVBench 等多项都拿了第一。\n\n但官方自己也承认短板:HLE 总分 43.6,落后 Fable 5 的 53.3 与 GPT-5.6 Sol 的 47.2;OSWorld 2.0 二进制得分 19.4,低于 Fable 5 的 66.1;Toolathlon Verified 72.5,也低于 Fable 5 的 77.9。换句话说,**它在 coding 与视觉侧追平甚至超越,在通用推理与超长任务规划上仍落后 Claude 与 GPT 的顶级版本**。\n\n## 长上下文能力:1M 是默认,工具型 agent 已经上手\n\n1M 上下文窗口是 Qwen3.8 全系的标配,官方在最佳实践里建议把内部推理的输出上限放到 262,144 token、最终回答放到 131,072 token,以便在长链工具调用中保持完整推理轨迹。\n\nQwen3.8 同步支持 `reasoning_effort`(xhigh \u002F medium \u002F low)控制思考深度,`preserve_thinking` 默认开启,让跨轮思考内容可以持续累积——这正是 agent 场景里\"不能每轮重置思维链\"的关键。MRCR v2 256K(8-needle)拿到 92.9,LongBench v2 拿到 66.3,与 GPT-5.6 Sol 的 93.8 \u002F 67.1 几乎咬平。\n\nAPI 上同时走 OpenAI Chat Completions、OpenAI Responses、Anthropic Messages 三种协议,Claude Code、Codex、Qoder CLI、Qwen Code、OpenClaw 都能直接换 base_url 接进来。\n\n## 三个长程演示:连续 16 天的自动编程\n\n发布博客里最让同行坐不住的,是官方做的三场\"长程自主\"演示。\n\n- **10+ 天跑出一个自演化工程框架**。Qwen3.8-Max 被要求从零创建 oh-my-cli 项目,16 天连续自主运行,累计 265 次 commit、127 个 PR、151 个 issue,GitHub 仓库全程公开。\n- **复现论文再改进**。从一篇 arXiv 论文出发,5 天连续工作、约 125 小时、约 7,600 行代码、33 轮 GPU 训练、1,100+ 次行动,先把原论文 6 个主要结论复现出来,再做 4 轮共 18 个自提改进思路,最终在 AIME24 上比原论文方法再高 2.71 个百分点。\n- **24 小时击败 526 支人类队伍**。在天池 WWW2025 多模态对话意图识别赛上,以 45 次提交、最终 0.853 的准确率超过 458 \u002F 526(87%)的人类参赛队。\n\n这三件事都不是 benchmark 上的单项冠军,而是把\"模型能不能在一段时间里自己扛住一个目标\"的工程性问题摆到台面上。\n\n## 工作侧:覆盖几百个高价值职业\n\n另一块被官方重点强调的是\"work\"。模型在几百个职业的真实工作流上做了 stress test:从几百份合同里挑出 1,284 个相关条款(过去法务助理团队一周的工作量);一次产出 8 屏交互原型、零修改迭代(常规 UI\u002FUX 工作流要 3–5 轮);从 100+ 份食材供应简报一次性生成 26 道菜的菜单,食材成本锁定在 33.8%;从单套图纸在浏览器里复刻一栋 30 层写字楼抗震结构模型;把一张 2D 康复评估表升级为 3D 交互演示;把每位球员 ~8,400 次进攻 \u002F 防守回合解析成可直接用的战术画像。\n\nQwen3.8-Max 还展示了一条端到端的 ETF 轮动量化策略研发流程:从一行任务描述出发,系统自动规划工作流,自主完成因子构建、多轮贪心迭代、动态分析回测结果,并在发现\"设计期与验证期指标背离\"等过拟合信号时自动触发剪枝。另一个广度方向上,从 6 段短描述(动量、价值、质量、投资、低风险、情绪)派生出 50 个研究方向、~330 个子 agent、~6,000 次回测,所选因子的超额 Sharpe 落在 0.64–1.48。\n\n## 短板与下一步\n\n官方自己点名的局限也很坦率:benchmark 只覆盖 10 类 alignment failure,**无法覆盖那些尚未被定义、罕见或新涌现的对齐失败**;强化学习与生产级 RLHF 后训练可能让这些收益消退;AAR 只拒绝在 MMLU \u002F GSM8K \u002F IFEval 上的能力退化,其它能力维度未必测到。\n\n接下来三个研究方向:罕见风险预测器(用历史模型代际数据预测下一代会出现什么对齐失败)、研究结果预测器(在昂贵的长时间实验跑之前先预测是否有效)、更好的控制脚手架(让自动化对齐研究自身更可监控)。\n\n## 行业含义\n\n把 Max 级权重开源这一步,直接改写了\"开源 vs 闭源\"的边界:开源端第一次跨入 2T+ 总参数 \u002F 95B 激活的体量,而 Gated DeltaNet + QSA 的混合架构又把长上下文的算力与显存成本压到了一个新基线。对 Claude \u002F GPT 的闭源阵营来说,**真正的压力不是某一项 benchmark 被追平,而是开源模型的\"长程自主 + 工作流覆盖\"开始有了工程化样本**。\n\n对开发者,意味着 1M 默认窗口 + 三协议 API + reasoning_effort + preserve_thinking + 多 agent harness(Claude Code \u002F Codex \u002F Qoder \u002F OpenClaw)直接打通——不用再为\"agent 用什么底座\"纠结。\n\nQwen 团队在引用块里直接点了一句:\"Qwen3.8-Max is our most capable model to date, and the first open-weight model at Max scale\"。这句话既是产品口号,也是 2026 年下半年开源大模型的现实坐标。\n\n原文与基准表:[Qwen3.8-Max: A New Bar for Coding and Cowork](https:\u002F\u002Fqwen.ai\u002Fblog?id=qwen3.8)。","https:\u002F\u002Fqwen.ai\u002Fblog?id=qwen3.8","c36a21ac-2a77-421b-9519-1e150695732a",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"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},"47bcf3fb-8abd-4428-8b85-d5eda981b076","en","Qwen3.8-Max Open-Sources 2.4T Weights: Gated DeltaNet Cuts Long-Context Cost 87%","Alibaba's Qwen team officially open-sourced Qwen3.8-Max weights on August 3, totaling 2.4T parameters with 95B active; the sister Flash-Next variant combines Gated DeltaNet with Qwen Sparse Attention, sharply lowering compute and memory cost on long contexts.","Alibaba's Qwen team officially released Qwen3.8-Max on August 3 — and for the first time, they committed to open-sourcing a Max-class model's weights: 2.4T total parameters, 95B active, large enough to cross the \"largest open-weight LLM\" line, with weights promised for Hugging Face and ModelScope \"next week\".\n\n## Not just another big model: an architectural swap\n\nCompared to the Qwen3.5-era Gated DeltaNet + sparse MoE hybrid, Qwen3.8 changes four things at the foundation. Flash-Next serves as this generation's architectural preview: 125B main model + 51B N-gram embeddings, with 6B parameters activated per token. The changes fall on four layers.\n\n- **Attention**: Gated DeltaNet (GDN) continues to compress history; the newly added **Qwen Sparse Attention (QSA)** uses a compressed lightweight indexer to pick salient context at micro-block granularity, cutting the cost of attention on long sequences.\n- **Residual stream**: **Gated Residual (GR)** widens the residual stream into 4 branches and controls reads\u002Fwrites through a dynamic gate, stabilizing cross-layer information flow and training stability.\n- **Embedding**: **N-gram Embedding** uses local context to look up tables that scale capacity at very little extra compute; the embedding table can be offloaded to host memory and overlapped with model computation via asynchronous prefetching.\n- **Optimizer**: the whole stack moves to the **Muon optimizer**, re-tuned around orthogonalization accuracy, the division of labor between Muon and AdamW, and the splitting of fused parameters; scaling laws are also re-fit for the new architecture.\n\nIn official framing, the position of this new architecture is unambiguous: Qwen3.8-Flash-Next plays the same role Qwen3-Next once played for Qwen3.5 — ship the architectural change first so the community can stress-test it, then stack the full Qwen4 family on top.\n\n## Benchmarks: head-on against GPT-5.6 Sol, Fable 5, Opus 4.8\n\nThe official comparison table covers 38 benchmarks. Highlights: Terminal Bench 2.1 — Qwen3.8-Max 86.6, close to GPT-5.6 Sol's 88.8, beating Opus 4.8 and Fable 5 at 84.6; PaperBench 93.0, beating GPT-5.6 Sol (90.5), Fable 5 (88.8), Opus 4.8 (80.3); AndroidBench 75.1, still behind Fable 5's 84.5 but ahead of GPT-5.6 Sol's 74.0.\n\nOn the vision side the gap is more dramatic: BabyVision 82.0 \u002F 91.3 (without \u002F with Code Interpreter), leaving Opus 4.8's 28.4 \u002F 81.2 far behind; HLE-VL (w\u002F Tools) 52.2, ahead of GPT-5.6 Sol's 51.2; ZeroBench Pass@5 24.0 \u002F 49.0, also the highest in the field. Across multimodal document and video benchmarks — OmniDocBench 1.5, Video MME (w\u002F Sub.), MLVU (M-Avg), LVBench — Qwen3.8-Max takes first place on multiple items.\n\nBut the team is upfront about the gaps: HLE 43.6 trails Fable 5's 53.3 and GPT-5.6 Sol's 47.2; OSWorld 2.0 binary score 19.4 is below Fable 5's 66.1; Toolathlon Verified 72.5 is also below Fable 5's 77.9. In short, **on coding and vision Qwen3.8-Max matches or surpasses the closed frontier; on general reasoning and ultra-long-horizon planning it still trails the top Claude and GPT variants**.\n\n## Long-context: 1M is default, agent tools already ship\n\nA 1M context window is standard across the Qwen3.8 line. The official best-practice guide recommends 262,144 tokens for internal reasoning output and 131,072 tokens for the final response within the 1M budget, so long tool-call chains can keep a complete reasoning trace.\n\nQwen3.8 also supports `reasoning_effort` (xhigh \u002F medium \u002F low) for explicit thinking-depth control and enables `preserve_thinking` by default, so the chain of thought accumulates across turns — exactly what agent scenarios need (\"don't reset reasoning every turn\"). MRCR v2 256K (8-needle) reaches 92.9, LongBench v2 66.3 — essentially neck-and-neck with GPT-5.6 Sol's 93.8 \u002F 67.1.\n\nOn the API side, Qwen3.8 serves OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages simultaneously, so Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw can drop in with a base_url swap.\n\n## Three long-horizon demos: 16 days of autonomous coding\n\nThe most uncomfortable part of the launch blog is the trio of \"long-horizon autonomy\" demos.\n\n- **A self-evolving engineering harness, 10+ days**. Asked to create the oh-my-cli project from scratch, Qwen3.8-Max ran autonomously for 16 consecutive days, accumulating 265 commits, 127 PRs, 151 issues — the GitHub repository is fully public.\n- **Reproduce a paper, then improve it**. Starting from a single arXiv paper, the model worked continuously for ~5 days — about 125 hours, ~7,600 lines of code, 33 rounds of GPU training, 1,100+ actions — first reproducing the paper's 6 main findings, then running 4 rounds of self-proposed improvements (18 ideas in total) and beating the paper's method by +2.71 points on AIME24.\n- **Beat 526 human teams in 24 hours**. On the Tianchi WWW2025 Multimodal Dialogue Intent Recognition Challenge, with 45 submissions the model climbed from 0.60 to a final 0.853 accuracy, beating 458 of 526 (87%) human teams.\n\nNone of these is a single benchmark crown. Together they put the engineering question \"can the model carry an open-ended goal on its own for a stretch of time?\" on the table.\n\n## Work side: covering hundreds of high-value professions\n\nThe other piece the team emphasizes is \"work\" — the messy, multi-step, tool-heavy tasks that drive economic value. The model was stress-tested across hundreds of real professional workflows: surfacing 1,284 relevant clauses across hundreds of contract documents in one pass (a paralegal team's week of collaborative work); producing an 8-screen interactive prototype with zero rounds of human revision (a conventional UI\u002FUX flow takes 3–5 rounds); generating a complete 26-dish menu from 100+ ingredient briefs, holding food cost at 33.8%; reconstructing a 30-story office tower's seismic structural model from a single drawing set, inspectable in browser; turning a 2D rehab assessment form into a 3D interactive demo; parsing ~8,400 possessions per player into ready-to-use tactical profiles.\n\nQwen3.8-Max also demonstrates an end-to-end ETF rotation quant strategy pipeline: starting from a one-line task description, the model autonomously plans the workflow, builds the data system, runs multi-round greedy iteration, and triggers pruning automatically when it detects design-vs-validation-period metric divergence (a classic overfitting signal). On the breadth side, 6 short descriptions (momentum, value, quality, investment, low-risk, sentiment) are decomposed into 50 research directions, ~330 sub-agents, and ~6,000 backtests, with selected factors achieving excess Sharpe ratios of 0.64–1.48.\n\n## Gaps and next steps\n\nThe team's own caveats are candid: the benchmark suite covers only 10 categories of alignment failures and **cannot cover unknown, rare, or newly emergent failures**; gains may not survive intensive production RLHF post-training; the AAR capability gate only rejects regressions on MMLU \u002F GSM8K \u002F IFEval, leaving other capability dimensions unmeasured.\n\nThree follow-up research directions are listed: a rare-risk forecaster (predict what alignment failures will appear in the next model generation from historical data), a research-outcome forecaster (predict whether a long-horizon experiment will work before running it), and stronger control scaffolding (so automated alignment research is more monitorable).\n\n## Industry implications\n\nPutting Max-class weights open-source redraws the open-vs-closed frontier: open weights cross the 2T+ total \u002F 95B active threshold for the first time, and the Gated DeltaNet + QSA hybrid pushes long-context compute and memory cost down to a new baseline. For Claude \u002F GPT, **the real pressure is not that any single benchmark has been matched, but that the open-source camp now has engineering samples for \"long-horizon autonomy + workflow coverage\"**.\n\nFor developers, this means 1M default context + three-protocol API + `reasoning_effort` + `preserve_thinking` + multi-agent harness support (Claude Code \u002F Codex \u002F Qoder \u002F OpenClaw) out of the box — no more hand-wringing over which base model to put under an agent.\n\nThe Qwen team put it bluntly in the citation block: \"Qwen3.8-Max is our most capable model to date, and the first open-weight model at Max scale.\" That's both a product tagline and the actual coordinate of the open-source LLM world in the second half of 2026.\n\nOriginal article and benchmark tables: [Qwen3.8-Max: A New Bar for Coding and Cowork](https:\u002F\u002Fqwen.ai\u002Fblog?id=qwen3.8).","qwen3-8-max-2-4t-open-weights-gated-deltanet","2026-08-30T03:00:00Z","2026-08-30T11:05:47.932431Z","2026-08-30T11:05:47.932442Z",true,"agent",162,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"d941056b-c2e7-42e5-965a-a982c20b1169","Qwen3.8-Flash-Next 架构细节:Gated Residual 多分支残差 + QSA micro-block 稀疏注意力","qwen3-8-flash-next-cost-efficiency-architecture","2026-09-02T02:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"40095b51-97b0-4fd4-9b1d-f636c970572e","阿里 Qwen 团队发布 Qwen3.8-Max:2.4 万亿参数 MoE 模型首度开放权重","qwen3-8-max-2-4t-moe-open-weights","2026-08-07T02:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"df01dae0-0940-4947-a019-31c57066132c","阿里千问 Qwen3.8 预览版上线:2.4T 参数,把开源旗舰抬到 Fable 5 同一档","qwen-3-8-preview-2-4t","2026-07-19T10:01:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"ea425005-49e7-477b-9f64-54361254c2d2","Qwen 开进驾驶场景:Qwen-Drive-1.0 保留 VLM 主干,外挂 BEV 感知与规划专家","qwen-drive-1-vlm-autonomous-driving","2026-09-02T19:35:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"741bd34c-7134-4e8e-ab45-4f53dc576a6b","腾讯 Hy4 登顶 9 月开源榜:79.87 分超 Qwen3.8 Max,Anthropic 包揽总榜前三","tencent-hy4-tops-open-source-benchlm-september","2026-09-01T17:10:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"932ae5e5-3552-4f9d-a6fe-26eedca0bb2b","蚂蚁首个金融增强模型开源在即:Ling-3.0-flash-Fin 押注投研 Agent","ant-ling-3-flash-fin-finance-llm","2026-08-31T19:15:00+00:00"]