本文来源:https://www.donews.com/news/detail/1/6772283.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
11月16日消息,阿里通义千问团队今日宣布推出新一代旗舰模型 Qwen4-Max,采用全新的 MoE(混合专家)架构,主干参数规模达到 2.4 万亿,原生支持视觉理解,并提供 100 万 Token 上下文窗口。
与此前 11 月 11 日先行开放的限时预览版不同,此次发布标志着 Qwen4-Max 作为正式版本接入千问 AI 平台,开发者可通过 Token Plan 或 API 调用。
Qwen4-Max 采用稀疏 MoE 架构,围绕 Attention、Residual 和 Embedding 三个层面做了系统性调整。每四层中三层使用 Gated DeltaNet 将历史信息压缩为固定状态,剩余一层以 Qwen Sparse Attention 按 block 粒度选择相关上下文,以降低长序列的注意力计算开销;Gated Residual 将残差流扩展为四条并行分支,增强跨层信息传递;N-gram Embedding 则在第二层以查表方式提供局部 token 模式表示,以较低额外算力扩展模型容量。
官方称,新模型在提升复杂工作任务处理能力的同时,能够减少 Token 消耗并降低任务成本。
根据 Qwen 官方博客介绍,Qwen4-Max 面向办公生产力、创意创作、软件开发和科学研究等场景,重点提升专业工具使用能力、长周期任务执行能力以及最终交付质量。模型能够处理 3D 建模、游戏开发、工业制造、金融、法律、医疗和前沿科学研究等任务。
效率是此次升级的显性指标。Qwen4-Max 在 Work360 行业基准上取得 52.69 分,前代 Qwen3.8-Max 为 47.74 分。
在同等性能下,Xhigh 和 High 两档推理强度的 token 消耗分别比前代减少约 24% 和 35%;Terminal-Bench 4.0 上的 token 消耗降幅则达到约 48% 和 49%。API 定价同步下调,尤其大幅降低了缓存命中价格,Work360 评测中单个任务的平均费用从 26.58 元降至 11.16 元,降幅 58%。
为了评估模型在真实工作场景中的表现,Qwen 团队还开发了 Work360 评测框架,用于模拟和分类复杂的专业工作任务。该框架覆盖 26 个一级领域、182 个二级子领域和 721 个三级专业方向,涉及多种行业和职业工作流程。
模型能力向长程任务执行倾斜。Qwen4-Max 支持百万 token 上下文窗口和原生视觉输入,在智能体场景中引入 Live Delivery 机制:先交付可用首版,再根据用户反馈逐版扩展内容与质量,同步开源了 Qwen-Live-Delivery-Plugin 插件。
官方展示的案例覆盖 Blender 三维建模、Unity 与 Unreal 引擎游戏开发、汽车产线规划与零件加工编程,以及金融投研、法律尽调和危重症临床辅助等专业场景。在汽车零件加工案例中,单个工件编程的人工投入从约 475 分钟降至 24 分钟。
Qwen4-Max 是 Qwen4 系列的首款模型。官方表示,系列中的小尺寸模型将陆续开源,沿袭同源技术路线,面向小模型更擅长的场景做取舍。
目前,Qwen4-Max API 已在 QwenCloud 上线,开发者可以通过 API 接入。此外,用户还可以通过 Qwen Studio、QwenWork 和 Qoder 体验新模型。
本文来源:https://www.donews.com/news/detail/1/6772283.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
On November 16 Alibaba's Qwen team announced Qwen4-Max, a new flagship model built on a fresh mixture-of-experts architecture with 2.4 trillion total parameters, native visual understanding and a one-million-token context window. Unlike the limited preview opened on November 11, this release makes Qwen4-Max a generally available version on the Qwen AI platform, accessible through Token Plan or API.
The sparse MoE design was reworked systematically at three levels — attention, residual connections and embeddings. Three out of every four layers use Gated DeltaNet to compress history into a fixed-size state, while the remaining layer uses Qwen Sparse Attention to select relevant context at block granularity. Qwen says the model improves handling of complex work tasks while consuming fewer tokens and lowering the cost per completed task.
Qwen4-Max targets office productivity, creative work, software development and scientific research, with emphasis on professional tool use, long-horizon task execution and final delivery quality, spanning 3D modelling, game development, industrial manufacturing, finance, law, medicine and frontier research. On the Work360 industry benchmark it scored 52.69, up from 47.74 for Qwen3.8-Max. At equal performance, token consumption at the Xhigh and High reasoning tiers fell about 24 and 35 percent respectively, and Terminal-Bench 4.0 saw reductions near 48 and 49 percent. API prices were cut alongside, especially the cache-hit rate, with average cost per task in the Work360 evaluation falling sharply.
To measure real-world behaviour the team built Work360, an evaluation framework that simulates and classifies complex professional work across 26 top-level domains, 182 second-level subdomains and 721 specialised directions. The model also leans into long-running agent work: with a million-token context and native visual input it introduces a Live Delivery mechanism that ships a usable first version, then expands content and quality iteratively from user feedback, accompanied by the open-source Qwen-Live-Delivery-Plugin.
