本次发布中，我们开放了 Qwen3.8-Flash-Next 的权重。这是一款多模态 MoE 模型，同时也是 Qwen4 所采用架构的早期预览。它所扮演的角色与 Qwen3-Next 之于 Qwen3.5 相同：当时引入的混合 Gated DeltaNet + Gated Attention 设计，此后已被广泛应用于 Qwen3.5、Qwen3.6、Qwen3.7 和 Qwen3.8 系列。
我们再次提前公开架构变更，以便社区在完整的 Qwen4 模型家族基于这些架构构建之前，能够先行审视。
Qwen3.8-Flash-Next 从四个方面——注意力、残差、嵌入和优化——对模型进行了系统性升级，在提升模型能力的同时，进一步优化了计算效率、模型容量和训练稳定性：
注意力：采用 GDN + QSA 混合架构。Gated DeltaNet（GDN）高效压缩历史信息；Qwen Sparse Attention（QSA）使用压缩后的轻量级索引器，在微块粒度上选择重要上下文，大幅降低长序列上的注意力计算成本。
残差：Gated Residual（GR）将残差流扩展为 4 个分支，并通过动态门控控制读写操作，强化了跨层信息流动和训练稳定性。
嵌入：N-gram Embedding 利用局部上下文查表，以极少的额外计算扩展模型容量；嵌入表可卸载至主机内存，并通过异步预取与模型计算重叠执行。
优化：采用 Muon 优化器，并围绕正交化精度、Muon 与 AdamW 之间的分工以及融合参数的拆分进行了精细调优，同时针对新架构重新拟合了缩放定律。
Qwen3.8-Flash-Next 拥有 125B 参数的主模型，并辅以额外的 51B N-gram 嵌入向量，每个 token 激活 6B 参数。与 Qwen3.7-Plus 相比，Qwen3.8-Flash-Next 大幅降低了训练和推理成本——训练量仅约为其 1/9，但在编程和办公任务上却展现出更优的能力。
它原生支持 262,144 个 token 的上下文，并可通过 YaRN 扩展至 1,000,000 个 token。有关 Qwen3.8-Flash-Next 架构、训练方法和实验分析的更多技术细节，请参阅我们 GitHub 仓库中的技术报告。
Qwen3.8-Flash-Next 的权重现已可在 Hugging Face 和 ModelScope 上获取。生产版本默认支持 1M 上下文并内置官方工具，以 Qwen3.8-Flash 的名称在 QwenCloud 上提供服务，定价为每百万输入 token 0.16 美元、每百万输出 token 0.47 美元（API 即将推出）。
语言#
Qwen3.8-Flash-NextQwen3.8-27BQwen3.7-PlusDeepSeek-V4-Flash-0731Claude-Opus-4.6 (Max)
参数量
125B27B397B284B--
激活参数量
6B27B17B13B--
N-gram 嵌入向量参数
51B--------
编码
智能体编码
DeepSWE 1.1
58.742.216.554.4--
智能体编码
SWE-bench Pro
62.561.755.856.053.4
多语言软件工程
SWE-bench Multilingual
81.073.875.8--77.5
仓库级代码生成
NL2Repo-Bench
48.142.341.154.247.6
智能体
长周期办公任务
CoWorkBench
73.970.765.145.168.2
专业工作任务
JobBench
55.733.427.641.336.6
前沿智能体任务
Agents' Last Exam
Pass@1
得分
Pass@1
得分
Pass@1
得分
Pass@1
得分
真实世界工具使用
Toolathlon Verified（Pass@1）
73.567.150.670.3--
通用
指令遵循
IFBench
81.379.579.179.262.5
科学推理
GPQA Diamond
91.789.290.390.891.3
多学科推理
HLE
35.930.834.733.840.0
竞赛级编程
LiveCodeBench v6
91.990.389.690.688.8
- DeepSWE 1.1：使用 Claude Code 和 mini-SWE-agent 两种测试框架进行评估，temp=1.0，top_p=0.95，上下文窗口 256K。我们报告两个框架中的最高得分；值得注意的是，Qwen3.8-Flash-Next 在 mini-SWE-agent 上表现最佳。
- SWE-bench Pro：除 Claude-Opus-4.6（Max）采用官方公布的分数外，所有模型均使用 Claude Code 测试框架进行评估，temp=1.0，top_p=0.95，上下文窗口 256K。有问题的任务已修正，所有基线模型均在优化后的基准上重新评估。
- SWE-bench Multilingual：使用 mini-SWE-agent 测试框架进行评估，temp=1.0，top_p=0.95，上下文窗口 256K。
- NL2Repo-Bench：使用 Claude Code 评测框架进行评估。为防止奖励黑客行为，我们禁用了试图访问特定仓库的 Bash 命令，例如 pip download、pip install 和 git clone。
- CoWorkBench：一个内部开发的协作基准测试，用于评估计算机科学、金融、法律、医学及其他生产力领域中涉及长周期办公与生产力任务的智能体能力。
- HLE：由 GPT-4o 进行评判。
- 每行中的最佳结果以粗体显示。
- 空白单元格（--）：表示分数尚不可用或不适用。
视觉语言#
Qwen3.8-Flash-NextQwen3.8-27BQwen3.7-PlusClaude-Opus-4.6 (Max)
智能体多模态智能
多模态工具调用
ClawEval-MM
Pass@3
平均
Pass@3
平均
Pass@3
平均
Pass@3
平均
应用复刻
RecreationBench
49.947.130.2--
移动端使用
AndroidWorld
84.581.981.062.0
电脑端使用
OSWorld 2.0
可视化网页开发
Vision2Web
64.062.942.1--
通用多模态智能
具身智能
ERQA
72.365.569.840.8
长视频理解
LVBench
76.672.476.263.0
真实世界感知
RealWorldQA
88.585.986.973.9
视觉数学问题求解
MathVision
无 CI
有 CI
无 CI
有 CI
无 CI
有 CI
无 CI
科学图表分析
CharXiv (RQ)
无 CI
有 CI
无 CI
有 CI
无 CI
有 CI
无 CI
- ClawEval-MM：分数以“pass@3 / 平均分”形式报告。Pass@3 衡量的是在三次试验中至少一次通过的百分比，平均分则是三次试验的平均得分。
- RecreationBench：一个内部的长时程应用复刻基准，用于评估跨五个平台的混合智能体能力——桌面端（Ubuntu、macOS、Windows）、移动端（Android）和网页端。
- OSWorld 2.0：分数以“binary / partial”形式报告。Binary 分数是获得完整任务奖励的任务百分比，而 partial 分数则汇总了所有任务中获得的局部奖励。
- Vision2Web：分数以前端、网页和网站类别的平均值报告，使用 Claude Code harness，并由 gpt-5.4-2026-03-05 评判。
- MathVision、CharXiv（RQ）：分数以“无 CI / 有 CI”形式报告。MathVision 中少量错误的地面真值标注在人工核验后已得到修正。我们模型的得分使用固定提示词评估，例如“请逐步推理，并将最终答案放在 \boxed{} 中。”对于其他模型，我们报告在有无 \boxed{} 格式的两次运行中得分较高的一次。
- 每行最佳结果以粗体显示。
- 空单元格（--）表示分数尚未公布或暂不适用。
要点：GDN + QSA，兼顾高效内存与精准检索#
传统的全量注意力（Full Attention）可以直接访问所有历史 token，但随着上下文不断变长，计算成本和 KV Cache 的内存访问成本都会大幅上升。
沿袭 Qwen3.5 中引入的架构设计，Qwen3.8-Flash-Next 采用了 GDN [1] + 注意力混合架构：每四层中有三层使用 Gated DeltaNet（GDN），将历史信息持续压缩到固定大小的状态中，其余一层则使用全局注意力（global Attention），在整个上下文中精准检索信息。
对于全局注意力，我们进一步引入了 Qwen 稀疏注意力（Qwen Sparse Attention，QSA）。稀疏注意力通过只关注重要上下文来降低长序列的计算量。然而，现有的方法（如 DSA [2]）仍然依赖 token 级索引器来定位重要位置；随着上下文不断增长，索引器本身也会成为不可忽视的计算开销来源。
QSA 进一步压缩了这一过程：一个轻量级索引器先将序列聚合为微块，在块级别估计上下文重要性，然后选择最相关的区域进行 Attention。这不仅降低了 Attention 本身的成本，也降低了识别重要上下文所需的索引开销。与跨层共享索引的方法 [3] 相比，QSA 在每一层内独立进行序列压缩，减少了对跨层 Attention 相似性的依赖，使其特别适合 GDN 与 Attention 层交错排列的 Hybrid 架构。
简而言之：GDN 高效地“记忆”，而 QSA 精准地“检索”。
在 1M token 规模下，QSA 的 Attention Kernel 在 Prefill 和 Decode 阶段分别实现了最高 7.6 倍和 4.9 倍的加速。在代表高缓存复用在线服务场景的实验设置中（90% 的 Prefix Cache 命中率），Qwen3.8-Flash-Next 在 1M-token 上下文长度下实现了 Qwen3.7-Plus 8.6 倍的 Prefill 吞吐量。
门控残差：为信息流动提供更多路径#
在传统的 Transformer 中，所有层都持续从同一个 Residual Stream 读取和写入。随着网络不断加深，早期特征会与后期信息反复混合，使得重要信号更有可能被逐渐稀释。
门控残差（GR）可以看作两种思路的结合：它沿用了 Hyper-Connection [4] 的做法，将残差流拓宽为多个分支，同时把 GatedNorm [5] 的逐元素动态门控引入残差读取过程。原本单一的残差流被扩展为四个并行分支，使模型能够根据当前内容动态决定从每个分支读取多少信息、向每个分支写回多少信息。
这可以理解为将单一信息通道扩展为多条并行通路：部分分支负责局部信息流动，而另一些分支则将早期信息直接保留并深入传递到网络深层。实证分析还表明，其中一个分支会自然涌现为一条长程通路，将第一个 Attention 层与中后段的大多数层连接起来。
GR 还进一步简化了 Hyper-Connection。一旦读取和写入操作具备足够的表达能力，额外的分支混合便不再带来显著收益，可以直接移除，从而降低内存访问开销和不稳定因素。Gate 还能有效抑制激活值离群点，提升训练稳定性。此外，Residual State 支持 FP8 存储，进一步减少了内存访问开销。
N-gram Embedding：以低成本扩展模型容量#
受 Gemma 3n 中的 Per-Layer Embedding 以及 DeepSeek Engram [6] 等工作的启发，我们进一步引入 N-gram Embedding，使模型容量能够超越 Transformer 主干网络的参数规模。
标准 Embedding 基于单个 token 进行查找。N-gram Embedding 则利用当前 token 及其前面若干 token 构成的局部上下文进行查找，为常见短语和局部模式提供额外的表示。
其关键优势在于，它可以在几乎不增加每个 token 计算量的情况下，引入大量参数。
Qwen3.8-Flash-Next 额外引入了 51B 的 N-gram Embedding 参数。由于查找位置可以预先确定，这些参数可以存储在主机内存中，并在模型计算的同时异步预取，无需永久占用 GPU 内存。
最终模型仅在网络靠近输入端的位置使用一个 N-gram Embedding 层，以相对较低的额外成本，有效地增加了一个大规模的“局部模式记忆”。
优化：架构与优化的协同设计#
Qwen3.8-Flash-Next 使用 Muon 优化器 [7] 进行训练，并围绕将 Muon 应用于大规模模型训练的三个方面进行了进一步改进：正交化精度、Muon 与 AdamW 之间的参数分配，以及融合参数矩阵的切分。
对于真正充当二维线性映射的参数，例如 Attention、GDN 和 MoE Experts 中的主要权重，我们使用 Muon。Embeddings、MoE Router 以及 GR 中的低秩参数继续使用 AdamW。对于在实现中融合的 QKV、SwiGLU 和 GDN 投影，我们首先根据它们所代表的独立线性变换进行拆分，然后分别进行正交化。
针对新架构和优化器，我们重新拟合了 Scaling Law。结果表明，模型可以稳定地使用更大的学习率和批大小，进一步提升了收敛效率和大规模并行训练吞吐量。
我们还发现，大规模模型训练中的常见做法——批大小预热（Batch Size Warmup）已不再必要：从小批量逐步增加到目标批量并不能改善最终结果，反而需要多出 18.8% 的优化器步数。因此，在最终的训练方案中，我们直接以目标批大小开始训练。
其他架构优化#
其余组件遵循 Qwen3-Next 中确立、并在 Qwen3.5–Qwen3.8 系列中不断完善的设计。
超稀疏 MoE：通过全局负载均衡 [8]，在保持激活专家数量不变的情况下增加专家总参数量，可以稳定地降低训练损失。因此，Qwen3.8-Flash-Next 使用大型专家池，每个 token 路由到少量专家，并配有一个共享专家。
多 token 预测：MTP 模块采用多步训练，保持训练与推理之间的一致性，从而提升投机解码在真实场景中的接受率，同时增强主干网络的性能。其全注意力层同样替换为 QSA。
训练稳定性：保留了对归一化权重施加权重衰减的零中心 RMSNorm、注意力输出门控机制 [9]，以及归一化的 MoE 路由器初始化。这些设计使小规模消融实验更加可靠，并有助于大规模训练平稳进行。
基座模型性能#
我们将 Qwen3.8-Flash-Next-Base 与 Qwen3.8-27B 和 Qwen3.7-Plus 的基座模型进行对比。
Qwen3.8-Flash-Next-BaseQwen3.8-27B-BaseQwen3.7-Plus-Base
参数量
125B27B397B
激活参数量
6B27B17B
N-gram 嵌入参数量
51B----
通用任务
MMLU
90.3687.5190.43
MMLU-Redux
90.6887.2691.47
MMLU-Pro
73.2368.6070.90
SuperGPQA
51.3644.8648.42
BBH
90.8789.5689.41
数学与 STEM 任务
GPQA
51.4245.0151.52
GSM8K
93.2993.1892.95
MATH
72.7860.5474.38
编程任务
EvalPlus
78.7676.0578.06
MultiPL-E
79.0974.5081.68
SWEBench-Pretrain
50.9941.6649.24
多语言任务
MGSM
89.3386.3785.42
MMMLU
84.8679.7484.53
INCLUDE
78.4074.3778.90
- 每一行中的最佳结果以粗体显示。
- 空白单元格（--）：分数尚未公布或暂不适用。
在 6B 激活参数下，Qwen3.8-Flash-Next-Base 在 14 项基准测试中的 8 项上取得了最佳结果，包括 MMLU-Pro、SuperGPQA、BBH、GSM8K、EvalPlus、SWEBench-Pretrain、MGSM 和 MMMLU，并且在 MMLU、MMLU-Redux、GPQA、MATH 和 MultiPL-E 上保持与 Qwen3.7-Plus-Base 接近的水平。
51B 的 N-gram 嵌入参数采用确定性寻址，不进入逐 token 的矩阵乘法预算。
使用 Qwen3.8-Flash-Next 进行开发#
Qwen3.8-Flash-Next 以开放权重模型的形式在 HuggingFace 和 ModelScope 上提供，并可通过 QwenCloud 上的官方托管 API 使用。该模型旨在平衡能力、延迟和成本，非常适合高吞吐量应用、工具驱动的工作流以及编码与协作助手。下面，您可以了解如何调用 QwenCloud API，并将 Qwen3.8-Flash-Next 集成到智能体系统和编码助手中。
即将推出。该 API 尚未上线——将在本文发布后不久启用。本节中的示例届时即可使用。
Qwen3.8-Flash-Next 可通过 API 使用：
QwenCloud#
在 QwenCloud 上，该模型以 qwen3.8-flash 的名称提供服务。QwenCloud 支持行业标准协议，包括兼容 OpenAI 的 Chat Completions 和 Responses API，以及兼容 Anthropic 的接口。
""" 环境变量： DASHSCOPE_API_KEY：您的 API 密钥，来自 https://home.qwencloud.com/ DASHSCOPE_BASE_URL：（可选）兼容模式 API 的基础 URL。
- 北京：https://dashscope.aliyuncs.com/compatible-mode/v1
- 新加坡：https://dashscope-intl.aliyuncs.com/compatible-mode/v1
- 美国（弗吉尼亚）：https://dashscope-us.aliyuncs.com/compatible-mode/v1 """ from openai import OpenAI import os api_key = os.environ.get("DASHSCOPE_API_KEY") if not api_key: raise ValueError( "DASHSCOPE_API_KEY is required. " "Set it via: export DASHSCOPE_API_KEY='your-api-key'" ) client = OpenAI( api_key=api_key, base_url=os.environ.get( "DASHSCOPE_BASE_URL", "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", ), ) messages = [{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}] completion = client.chat.completions.create( model="qwen3.8-flash", messages=messages, extra_body={ "enable_thinking": True,
"preserve_thinking": True,
}, reasoning_effort="xhigh", # supported levels are xhigh, medium, and low stream=True, ) reasoning_content = "" answer_content = "" is_answering = False print("\n" + "=" * 20 + "Reasoning" + "=" * 20 + "\n") for chunk in completion: if not chunk.choices: print("\nUsage:") print(chunk.usage) continue delta = chunk.choices[0].delta if hasattr(delta, "reasoning_content") and delta.reasoning_content is not None: if not is_answering: print(delta.reasoning_content, end="", flush=True) reasoning_content += delta.reasoning_content if hasattr(delta, "content") and delta.content: if not is_answering: print("\n" + "=" * 20 + "Answer" + "=" * 20 + "\n") is_answering = True print(delta.content, end="", flush=True) answer_content += delta.content
智能体框架与编程助手#
Qwen3.8-Flash-Next 可与主流智能体框架和编程助手无缝集成：
QwenWork（即将推出）#
QwenWork 是阿里巴巴旗舰级 AI 生产力平台，旨在帮助个人和企业自动化日常任务并提升运营效率。
我们很高兴地宣布，QwenWork 已集成 Qwen3.8-Flash-Next，为其新推出的“标准”模式提供支持，借助该模型的前沿能力，提供流畅且高性价比的体验，为职场 AI 智能体树立新标杆。
更多详情请参阅官方文档！
Claude Code#
npm install -g @anthropic-ai/claude-code export ANTHROPIC_MODEL="qwen3.8-flash" export ANTHROPIC_SMALL_FAST_MODEL="qwen3.8-flash" export ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropic export ANTHROPIC_AUTH_TOKEN= claude
Codex#
Qwen API 支持 OpenAI Responses 协议，因此可与 Codex 配合使用：
在 ~/.codex/model-catalog.local.json 中
{ "models": [ { "slug": "qwen3.8-flash", "display_name": "qwen3.8-flash", "description": "QwenCloud: Qwen3.8-Flash", "default_reasoning_level": "xhigh", "supported_reasoning_levels": [ { "effort": "low", "description": "快速响应，推理较轻" }, { "effort": "medium", "description": "针对复杂问题提供更深入的推理" }, { "effort": "xhigh", "description": "针对复杂问题提供超高深度推理" } ], "context_window": 1000000, "effective_context_window_percent": 95, "supports_parallel_tool_calls": true, "supports_image_detail_original": true, "input_modalities": ["text", "image"], "shell_type": "default", "visibility": "list", "supported_in_api": true, "priority": 1, "base_instructions": "", "support_verbosity": false, "supports_reasoning_summaries": false, "experimental_supported_tools": [], "truncation_policy": { "mode": "bytes", "limit": 10000 } } ] }
在 ~/.codex/config.toml 中
model_catalog_json = "~/.codex/model-catalog.local.json" model_provider = "QwenCloud" model = "qwen3.8-flash" [model_providers.QwenCloud] name = "QwenCloud" base_url = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1" env_key = "OPENAI_API_KEY" wire_api = "responses"
npm install -g @openai/codex export OPENAI_API_KEY= codex
Qoder CLI#
Qoder 与 Qwen 协同进化，赋能智能体编程：
curl -fsSL https://qoder.com/install | bash qoder
Qwen Code#
npm install -g @qwen-code/qwen-code@latest qwen
OpenClaw#
通过 QwenCloud 连接 OpenClaw：
curl -fsSL https://openclaw.ai/install.sh | bash export DASHSCOPE_API_KEY= openclaw dashboard
{ "models": { "mode": "merge", "providers": { "qwencloud": { "baseUrl": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", "apiKey": "DASHSCOPE_API_KEY", "api": "openai-completions", "models": [ { "id": "qwen3.8-flash", "name": "qwen3.8-flash", "reasoning": true, "input": ["text", "image"], "contextWindow": 1000000, "maxTokens": 65536 } ] } } }, "agents": { "defaults": { "model": { "primary": "qwencloud/qwen3.8-flash" } } } }
Qwen3.8-Flash-Next 在 Qwen3-Next 引入的混合架构基础上，沿四个方向进行了扩展：注意力、残差、嵌入向量和优化。QSA 将序列压缩到每一层内的微块中，在长上下文场景下同时降低注意力计算成本和索引成本，同时保持精确检索。
Gated Residual 将残差流拓宽为若干并行分支，并通过逐元素、数据相关的门控机制控制读写，以可忽略不计的算术开销改善跨层信息流动和训练稳定性；残差状态还可额外以 FP8 格式保存，进一步减少内存流量。N-gram 嵌入向量通过确定性寻址的查找内存扩展容量，该内存可以以可忽略不计的每 token 计算量进行扩展，并可卸载到主机内存。
在优化方面，Muon 被用作主优化器，正交化精度、参数分配和融合矩阵切分是决定性的实现选择，并且针对新架构重新拟合了缩放定律。
我们提前发布这些权重，以便社区能够像对待 Qwen3-Next 那样独立评估该架构，我们将继续对其进行优化，朝着 Qwen4 迈进。
@techreport{qwen2026design, title = {On the Design of {Qwen3.8-Next} Architecture: Evaluation, Efficiency, and Training Stability}, author = {{Qwen Team}}, institution = {Alibaba Group}, month = {August}, year = {2026} } @misc{qwen3.8flashnext, title = {{Qwen3.8-Flash-Next}: A New Architecture, Towards Ultimate Cost-Efficiency}, author = {{Qwen Team}}, month = {August}, year = {2026}, url = {https://qwen.ai/blog?id=qwen3.8-flash-next} }
参考文献#
[1] 门控 Delta 网络：用 Delta 规则改进 Mamba2
[2] DeepSeek-V3.2：推动开放大语言模型的前沿
[3] IndexCache：通过跨层索引复用加速稀疏注意力
[4] 超连接（Hyper-Connections）
[5] 注意力与残差汇（Residual Sinks）的统一视角：离群值驱动的重缩放对 Transformer 训练至关重要
[6] 通过可扩展查找实现条件记忆：大语言模型稀疏性的新维度
[7] Muon：面向神经网络隐藏层的优化器
[8] 细节中的魔鬼：论如何实现负载均衡损失以训练专业化混合专家模型
In this release we are opening the weights of Qwen3.8-Flash-Next, a multimodal MoE model that also serves as an early preview of the architecture used in Qwen4. It plays the same role that Qwen3-Next played for Qwen3.5: the hybrid Gated DeltaNet + Gated Attention design introduced at that time has since been used across the Qwen3.5, Qwen3.6, Qwen3.7 and Qwen3.8 series. We are again releasing the architectural changes early, so that the community can examine them before the full Qwen4 model family is built on top of them.
Qwen3.8-Flash-Next upgrades the model systematically along four aspects — attention, residual, embedding and optimization — improving model capability while further optimizing computational efficiency, model capacity and training stability:
Attention: A GDN + QSA hybrid architecture. Gated DeltaNet (GDN) compresses the history efficiently; Qwen Sparse Attention (QSA) uses a compressed lightweight indexer to select the important context at micro-block granularity, substantially reducing the cost of attention on long sequences.
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.
Embedding: N-gram Embedding looks up a table using the local context to scale model capacity with very little extra computation; the embedding table can be offloaded to host memory and overlapped with model computation through asynchronous prefetching.
Optimization: The Muon optimizer is used, 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.
Qwen3.8-Flash-Next features a 125B-parameter main model, supplemented by an additional 51B N-gram embeddings, with 6B parameters activated per token. Compared with Qwen3.7-Plus, Qwen3.8-Flash-Next substantially reduces both training and inference cost — training takes only about 1/9 as much, yet it delivers superior capabilities in coding and office tasks.
It natively supports 262,144 tokens of context and is extensible to 1,000,000 tokens with YaRN. For more technical details on the architecture, training methodology, and experimental analysis of Qwen3.8-Flash-Next, please refer to the technical report in our GitHub repository.
Qwen3.8-Flash-Next weights are now available on Hugging Face and ModelScope. The production version, with 1M context by default and official built-in tools, is served as Qwen3.8-Flash on QwenCloud, priced at 0.16 USD per million input tokens and 0.47 USD per million output tokens (API coming soon).
Language#
Qwen3.8-Flash-NextQwen3.8-27BQwen3.7-PlusDeepSeek-V4-Flash-0731Claude-Opus-4.6 (Max)
Params
125B27B397B284B--
Activated params
6B27B17B13B--
N-gram embedding params
51B--------
Coding
Agentic coding
DeepSWE 1.1
58.742.216.554.4--
Agentic coding
SWE-bench Pro
62.561.755.856.053.4
Multilingual software engineering
SWE-bench Multilingual
81.073.875.8--77.5
Repo-level code generation
NL2Repo-Bench
48.142.341.154.247.6
Agent
Long-horizon office work
CoWorkBench
73.970.765.145.168.2
Professional job tasks
JobBench
55.733.427.641.336.6
Frontier agentic tasks
Agents' Last Exam
Pass@1
Score
Pass@1
Score
Pass@1
Score
Pass@1
Score
Real-world tool use
Toolathlon Verified (Pass@1)
73.567.150.670.3--
General
Instruction following
IFBench
81.379.579.179.262.5
Scientific reasoning
GPQA Diamond
91.789.290.390.891.3
Multidisciplinary reasoning
HLE
35.930.834.733.840.0
Competitive coding
LiveCodeBench v6
91.990.389.690.688.8
- DeepSWE 1.1: evaluated with the Claude Code and mini-SWE-agent harnesses, temp=1.0, top_p=0.95, 256K context window. We report the highest score across the two harnesses; notably, Qwen3.8-Flash-Next performs best on mini-SWE-agent.
- SWE-bench Pro: except for Claude-Opus-4.6 (Max), for which we report the officially published score, all models are evaluated with the Claude Code harness, temp=1.0, top_p=0.95, 256K context window. Problematic tasks were corrected and all baseline models were re-evaluated on the refined benchmark.
- SWE-bench Multilingual: evaluated with the mini-SWE-agent harness, temp=1.0, top_p=0.95, 256K context window.
- NL2Repo-Bench: evaluated with the Claude Code harness. To prevent reward hacking, we disable Bash commands that attempt to access the specific repository, such as pip download, pip install and git clone.
- CoWorkBench: an in-house cowork benchmark for evaluating long-horizon office and productivity agent tasks across computer science, finance, law, medical and other productivity domains.
- HLE: judged by GPT-4o.
- The best result in each row is shown in bold.
- Empty cells (--): scores are not yet available or are not applicable.
Vision Language#
Qwen3.8-Flash-NextQwen3.8-27BQwen3.7-PlusClaude-Opus-4.6 (Max)
Agentic Multimodal Intelligence
Multimodal tool use
ClawEval-MM
Pass@3
Average
Pass@3
Average
Pass@3
Average
Pass@3
Average
Application recreation
RecreationBench
49.947.130.2--
Mobile use
AndroidWorld
84.581.981.062.0
Computer use
OSWorld 2.0
Visual web development
Vision2Web
64.062.942.1--
General Multimodal Intelligence
Embodied intelligence
ERQA
72.365.569.840.8
Long video understanding
LVBench
76.672.476.263.0
Real-world perception
RealWorldQA
88.585.986.973.9
Visual math problem solving
MathVision
Without CI
With CI
Without CI
With CI
Without CI
With CI
Without CI
Scientific chart analysis
CharXiv (RQ)
Without CI
With CI
Without CI
With CI
Without CI
With CI
Without CI
- ClawEval-MM: scores are reported as "pass@3 / average score". Pass@3 measures the percentage passed in at least one of three trials, and the average score is the mean score across the three trials.
- RecreationBench: an in-house long-horizon application-recreation benchmark for evaluating hybrid-agent abilities spanning five platforms — desktop (Ubuntu, macOS, Windows), mobile (Android) and web.
- OSWorld 2.0: scores are reported as "binary / partial". The binary score is the percentage of tasks that receive the full task reward, while the partial score aggregates the partial rewards obtained across all tasks.
- Vision2Web: scores are reported as the average over the frontend, webpage and website categories, using the Claude Code harness and judged by gpt-5.4-2026-03-05.
- MathVision, CharXiv (RQ): scores are reported as "without CI / with CI". A small number of incorrect ground-truth annotations in MathVision were corrected after manual verification. Our model's score is evaluated using a fixed prompt, e.g. "Please reason step by step, and put your final answer within \boxed{}." For other models, we report the higher score between runs with and without the \boxed{} formatting.
- The best result in each row is shown in bold.
- Empty cells (--) indicate scores not yet available or not applicable.
Attention: GDN + QSA for Efficient Memory and Precise Retrieval#
Traditional Full Attention provides direct access to all previous tokens, but as the context grows longer, both computation and KV Cache memory-access costs increase substantially.
Following the architecture design introduced in Qwen3.5, Qwen3.8-Flash-Next adopts a GDN [1] + Attention Hybrid architecture: three out of every four layers use Gated DeltaNet (GDN) to continuously compress historical information into a fixed-size state, while the remaining layer uses global Attention for precise retrieval of information across the full context.
For global Attention, we further introduce Qwen Sparse Attention (QSA). Sparse Attention reduces long-sequence computation by attending only to important context. However, existing approaches such as DSA [2] still rely on a token-level indexer to identify important positions; as the context grows, the indexer itself becomes a non-negligible source of computation.
QSA further compresses this process: a lightweight indexer first aggregates the sequence into micro-blocks, estimates context importance at the block level, and then selects the most relevant regions for Attention. This reduces not only the cost of Attention itself, but also the indexing overhead required to identify important context. Compared with approaches that share indices across layers [3], QSA performs sequence compression independently within each layer, reducing its dependence on cross-layer Attention similarity and making it particularly well suited to Hybrid architectures where GDN and Attention layers are interleaved.
Put simply: GDN efficiently “remembers,” while QSA precisely “retrieves.”
At 1M tokens, QSA’s Attention Kernel achieves up to 7.6× and 4.9× speedups in Prefill and Decode, respectively. In an experimental setup representative of online serving scenarios with high cache reuse (a 90% Prefix Cache hit rate), Qwen3.8-Flash-Next achieves 8.6× the Prefill throughput of Qwen3.7-Plus at a 1M-token context length.
Gated Residual: More Paths for Information Flow#
In a traditional Transformer, all layers continuously read from and write to the same Residual Stream. As the network becomes deeper, early features are repeatedly mixed with later information, making important signals more likely to be gradually diluted.
Gated Residual (GR) can be viewed as a combination of two ideas: it follows Hyper-Connection [4] in widening the residual stream into multiple branches, while incorporating the element-wise dynamic gating of GatedNorm [5] into the residual read. The original single residual stream is expanded into four parallel branches, allowing the model to dynamically determine how much information to read from each branch and how much to write back to each branch based on the current content.
This can be conceptualized as expanding a single information channel into multiple parallel pathways: some branches handle local information flow, while others preserve early information directly deep into the network layers. Empirical analysis also reveals that one of these branches naturally emerges as a long-range pathway connecting the first Attention layer to most of the middle and subsequent layers.
GR also further simplifies Hyper-Connection. Once the read and write operations are expressive enough, additional branch mixing yields no significant benefits and can thus be directly removed, thereby reducing memory access overhead and sources of instability. The Gate also effectively suppresses activation outliers and improves training stability. In addition, the Residual State supports FP8 storage, further reducing memory-access overhead.
N-gram Embedding: Expanding Model Capacity at Low Cost#
Inspired by Per-Layer Embedding in Gemma 3n and works such as DeepSeek Engram [6], we further introduce N-gram Embedding to scale model capacity beyond the parameters of the Transformer backbone.
A standard Embedding performs a lookup based on a single token. N-gram Embedding instead performs lookups using the local context formed by the current token and several preceding tokens, providing additional representations for common phrases and local patterns.
Its key advantage is that it can add a large number of parameters with almost no additional computation per token.
Qwen3.8-Flash-Next introduces an additional 51B N-gram Embedding parameters. Because lookup locations can be determined in advance, these parameters can be stored in Host Memory and asynchronously prefetched in parallel with model computation, without permanently occupying GPU memory.
The final model uses only a single N-gram Embedding layer near the beginning of the network, effectively adding a large-scale “local-pattern memory” at relatively low additional cost.
Optimization: Co-designing Architecture and Optimization#
Qwen3.8-Flash-Next is trained with the Muon Optimizer [7], with further improvements around three key aspects of applying Muon to large-scale model training: orthogonalization accuracy, parameter assignment between Muon and AdamW, and splitting fused parameter matrices.
For parameters that genuinely act as two-dimensional linear maps, such as the main weights in Attention, GDN, and MoE Experts, we use Muon. Embeddings, the MoE Router, and the low-rank parameters in GR continue to use AdamW. For QKV, SwiGLU, and GDN projections that are fused in the implementation, we first split them according to the independent linear transformations they represent, and then perform orthogonalization separately.
For the new architecture and Optimizer, we refit the Scaling Law. The results show that the model can stably use larger Learning Rates and Batch Sizes, further improving convergence efficiency and large-scale parallel training throughput.
We also find that Batch Size Warmup, a common practice in large-scale model training, is no longer necessary: gradually increasing from a small Batch to the target Batch does not improve the final result, but instead requires 18.8% more optimizer steps. In the final training Recipe, we therefore start directly with the target Batch Size.
Other Architecture Optimizations#
The remaining components follow the design established in Qwen3-Next and refined through the Qwen3.5–Qwen3.8 series.
Ultra-sparse MoE: With global load balancing [8], increasing total expert parameters while keeping the number of activated experts fixed steadily reduces training loss. Qwen3.8-Flash-Next therefore uses a large expert pool with a small number of routed experts per token, together with one shared expert.
Multi-Token Prediction: The MTP module is trained with multiple steps, maintaining consistency between training and inference and thereby improving the acceptance rate of speculative decoding in real scenarios, while also enhancing the performance of the backbone. Its full-attention layers are replaced with QSA as well.
Training stability: Zero-centered RMSNorm with weight decay applied to norm weights, the attention output gating mechanism [9], and normalized MoE router initialization are retained. These designs make small-scale ablations more reliable and help large-scale training run smoothly.
Base Model Performance#
We compare Qwen3.8-Flash-Next-Base with the base models of Qwen3.8-27B and Qwen3.7-Plus.
Qwen3.8-Flash-Next-BaseQwen3.8-27B-BaseQwen3.7-Plus-Base
Params
125B27B397B
Activated params
6B27B17B
N-gram embedding params
51B----
General tasks
MMLU
90.3687.5190.43
MMLU-Redux
90.6887.2691.47
MMLU-Pro
73.2368.6070.90
SuperGPQA
51.3644.8648.42
BBH
90.8789.5689.41
Math & STEM tasks
GPQA
51.4245.0151.52
GSM8K
93.2993.1892.95
MATH
72.7860.5474.38
Coding tasks
EvalPlus
78.7676.0578.06
MultiPL-E
79.0974.5081.68
SWEBench-Pretrain
50.9941.6649.24
Multilingual tasks
MGSM
89.3386.3785.42
MMMLU
84.8679.7484.53
INCLUDE
78.4074.3778.90
- The best result in each row is shown in bold.
- Empty cells (--): scores are not yet available or not applicable.
With 6B activated parameters, Qwen3.8-Flash-Next-Base achieves the best result on 8 of the 14 benchmarks, including MMLU-Pro, SuperGPQA, BBH, GSM8K, EvalPlus, SWEBench-Pretrain, MGSM and MMMLU, and remains close to Qwen3.7-Plus-Base on MMLU, MMLU-Redux, GPQA, MATH and MultiPL-E. The 51B N-gram embedding parameters are deterministically addressed and do not enter the per-token matrix-multiplication budget.
Develop with Qwen3.8-Flash-Next#
Qwen3.8-Flash-Next is available as an open-weight model on HuggingFace and ModelScope, with official managed APIs on QwenCloud. Designed to balance capability, latency, and cost, it is well suited for high-volume applications, tool-driven workflows, and coding & coworking assistants. In the following, you can explore how to call the QwenCloud API and integrate Qwen3.8-Flash-Next into agentic systems and coding assistants.
Coming soon. The API is not live yet — it will be enabled shortly after this post goes out. The examples in this section will work from then on.
Qwen3.8-Flash-Next is available via API:
QwenCloud#
On QwenCloud, the model is served under the name qwen3.8-flash. QwenCloud supports industry-standard protocols, including OpenAI-compatible Chat Completions and Responses APIs, alongside an Anthropic-compatible interface.
""" Environment variables: DASHSCOPE_API_KEY: Your API Key from https://home.qwencloud.com/ DASHSCOPE_BASE_URL: (optional) Base URL for compatible-mode API.
- Beijing: https://dashscope.aliyuncs.com/compatible-mode/v1
- Singapore: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
- US (Virginia): https://dashscope-us.aliyuncs.com/compatible-mode/v1 """ from openai import OpenAI import os api_key = os.environ.get("DASHSCOPE_API_KEY") if not api_key: raise ValueError( "DASHSCOPE_API_KEY is required. " "Set it via: export DASHSCOPE_API_KEY='your-api-key'" ) client = OpenAI( api_key=api_key, base_url=os.environ.get( "DASHSCOPE_BASE_URL", "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", ), ) messages = [{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}] completion = client.chat.completions.create( model="qwen3.8-flash", messages=messages, extra_body={ "enable_thinking": True,
"preserve_thinking": True,
}, reasoning_effort="xhigh", # supported levels are xhigh, medium, and low stream=True, ) reasoning_content = "" answer_content = "" is_answering = False print("\n" + "=" * 20 + "Reasoning" + "=" * 20 + "\n") for chunk in completion: if not chunk.choices: print("\nUsage:") print(chunk.usage) continue delta = chunk.choices[0].delta if hasattr(delta, "reasoning_content") and delta.reasoning_content is not None: if not is_answering: print(delta.reasoning_content, end="", flush=True) reasoning_content += delta.reasoning_content if hasattr(delta, "content") and delta.content: if not is_answering: print("\n" + "=" * 20 + "Answer" + "=" * 20 + "\n") is_answering = True print(delta.content, end="", flush=True) answer_content += delta.content
Agent Frameworks & Coding Assistants#
Qwen3.8-Flash-Next integrates seamlessly with popular agent frameworks and coding assistants:
QwenWork (Coming Soon) #
QwenWork is Alibaba’s flagship AI productivity platform, designed to help individuals and enterprises automate daily tasks and accelerate operational efficiency.
We are excited to share that QwenWork has integrated Qwen3.8-Flash-Next to power its newly launched “Standard” mode, leveraging the model’s cutting-edge capabilities to deliver a seamless, cost-effective experience that sets a new standard for AI agents in the workplace.
Learn more in the official documentation!
Claude Code#
npm install -g @anthropic-ai/claude-code export ANTHROPIC_MODEL="qwen3.8-flash" export ANTHROPIC_SMALL_FAST_MODEL="qwen3.8-flash" export ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropic export ANTHROPIC_AUTH_TOKEN= claude
Codex#
Qwen APIs support the OpenAI Responses protocol, enabling use with Codex:
In ~/.codex/model-catalog.local.json
{ "models": [ { "slug": "qwen3.8-flash", "display_name": "qwen3.8-flash", "description": "QwenCloud: Qwen3.8-Flash", "default_reasoning_level": "xhigh", "supported_reasoning_levels": [ { "effort": "low", "description": "Fast responses with lighter reasoning" }, { "effort": "medium", "description": "Greater reasoning depth for complex problems" }, { "effort": "xhigh", "description": "Extra high reasoning depth for complex problems" } ], "context_window": 1000000, "effective_context_window_percent": 95, "supports_parallel_tool_calls": true, "supports_image_detail_original": true, "input_modalities": ["text", "image"], "shell_type": "default", "visibility": "list", "supported_in_api": true, "priority": 1, "base_instructions": "", "support_verbosity": false, "supports_reasoning_summaries": false, "experimental_supported_tools": [], "truncation_policy": { "mode": "bytes", "limit": 10000 } } ] }
In ~/.codex/config.toml
model_catalog_json = "~/.codex/model-catalog.local.json" model_provider = "QwenCloud" model = "qwen3.8-flash" [model_providers.QwenCloud] name = "QwenCloud" base_url = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1" env_key = "OPENAI_API_KEY" wire_api = "responses"
npm install -g @openai/codex export OPENAI_API_KEY= codex
Qoder CLI#
Qoder co-evolves with Qwen for agentic coding:
curl -fsSL https://qoder.com/install | bash qoder
Qwen Code#
npm install -g @qwen-code/qwen-code@latest qwen
OpenClaw#
Connect to OpenClaw via QwenCloud:
curl -fsSL https://openclaw.ai/install.sh | bash export DASHSCOPE_API_KEY= openclaw dashboard
{ "models": { "mode": "merge", "providers": { "qwencloud": { "baseUrl": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", "apiKey": "DASHSCOPE_API_KEY", "api": "openai-completions", "models": [ { "id": "qwen3.8-flash", "name": "qwen3.8-flash", "reasoning": true, "input": ["text", "image"], "contextWindow": 1000000, "maxTokens": 65536 } ] } } }, "agents": { "defaults": { "model": { "primary": "qwencloud/qwen3.8-flash" } } } }
Qwen3.8-Flash-Next extends the hybrid architecture introduced in Qwen3-Next along four directions: attention, residual, embedding and optimization. QSA compresses the sequence into micro-blocks within each layer, reducing both the attention cost and the indexing cost at long context while keeping precise retrieval. Gated Residual widens the residual stream into several parallel branches and controls reads and writes with an elementwise, data-dependent gate, improving cross-layer information flow and training stability at negligible arithmetic cost; the residual state can additionally be kept in FP8, which further reduces memory traffic. N-gram embedding scales capacity through deterministically addressed lookup memory, which can be scaled with negligible per-token computation and offloaded to host memory. On the optimization side, Muon is used as the main optimizer, with orthogonalization accuracy, parameter assignment and fused-matrix splitting as the decisive implementation choices, and the scaling law refitted for the new architecture.
We release these weights early so that the architecture can be evaluated independently by the community, as we did with Qwen3-Next, and we will continue to refine it towards Qwen4.
@techreport{qwen2026design, title = {On the Design of {Qwen3.8-Next} Architecture: Evaluation, Efficiency, and Training Stability}, author = {{Qwen Team}}, institution = {Alibaba Group}, month = {August}, year = {2026} } @misc{qwen3.8flashnext, title = {{Qwen3.8-Flash-Next}: A New Architecture, Towards Ultimate Cost-Efficiency}, author = {{Qwen Team}}, month = {August}, year = {2026}, url = {https://qwen.ai/blog?id=qwen3.8-flash-next} }
References#
[1] Gated Delta Networks: Improving Mamba2 with Delta Rule
[2] DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
[3] IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse
[4] Hyper-Connections
[5] A Unified View of Attention and Residual Sinks: Outlier-Driven Rescaling is Essential for Transformer Training
[6] Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models
[7] Muon: An Optimizer for Hidden Layers in Neural Networks
[8] Demons in the Detail: On Implementing Load Balancing Loss for Training Specialized Mixture-of-Expert Models