• 亿邦公众号
    • APP下载

阿里Qwen4-Max发布 单任务成本降至前代三分之一

作者:亿邦AI 文章来源:亿邦动力

2026-11-17 16:11

作者:亿邦AI 文章来源:亿邦动力

2026-11-17 16:11

邦小白快读

EN
全文速览

总:Qwen4-Max最值得关注的是更强长文本与代理能力叠加显著降本,普通用户可直接用。

1.模型总参数2.4万亿但每次仅激活710亿,原生支持图像输入,上下文窗口100万token,适合专业长任务。

2.成本从每任务4.43美元降至1.67美元,降幅62%,Work360得分从47.74升至52.69;API每百万输入token收费1.50美元、输出4.50美元、缓存命中0.075美元。

3.实操上可在QwenCloud、QwenWork、Qoder和QwenStudio使用;Live Delivery先交付可用初版再逐步完善,用户可在任意节点介入并改变方向,插件已开源。

4.案例显示汽车领域单工件编程从约475分钟缩短到24分钟,金融场景可调用90个MCP工具生成可校验Excel和PowerPoint且数字可追溯。

总:也要看短板,Terminal-Bench 4.0得分48%低于Kimi K3.1 Pro的67%,FrontierCode 2.0为26%,选型要按具体任务。

总:对品牌商而言,文章提供的是阿里以低价高性能专业模型切入市场的产品定位与竞争策略,可用于判断定价、渠道和用户场景。

1.定价上,Qwen4-Max每项Intelligence Index任务成本3.27美元,约为前代10.02美元的三分之一,比Kimi K3.1 Pro的3.68美元低约11%;API输入1.50美元、输出4.50美元、缓存0.075美元,价格优势来自单价降25%、缓存便宜70%、耗token更少。

2.产品研发上,Qwen4架构混合三层压缩与第四层关键提取,加入四条并行连接和动态门控,并有3280亿参数N-gram嵌入层,长文本计算开销下降、训练稳定;原生图像和100万token面向工业、科研与金融。

3.用户行为上,模型围绕自主代理打造,Live Delivery先交付初版再逐步完善,用户可介入改方向,适合长任务;汽车、金融、游戏开发案例强调效率与可校验交付。

4.渠道上,已通过QwenCloud上线,在QwenWork、Qoder和QwenStudio可用,更小Qwen4计划以开源权重发布,形成云API加工具入口的分发。

总:品牌竞争上对标Opus 5.5、GPT-6和Kimi K3.1 Pro,指数43分落后Kimi的46分,但Agents Last Exam v1.1得分46.7为对比组最佳,需按场景塑造差异。

总:卖家的机会在于专业代理模型降本和工具生态开放,风险在于部分基准落后竞品,选型要按任务与合规。

1.机会上,单任务成本降62%,Intelligence Index每任务3.27美元低于Kimi的3.68美元;缓存命中便宜70%,API输入1.50美元、输出4.50美元,适合高频长任务。

2.可学习模式上,Live Delivery先交可用初版、后续迭代、用户可介入,降低长任务末段才暴露错误;金融案例调用90个MCP工具拉数据并输出可追溯Excel和PowerPoint;插件开源。

3.合作与扶持上,已通过QwenCloud上线,可在QwenWork、Qoder和QwenStudio使用;更小Qwen4计划开源权重;与Unity、Unreal Engine和Godot配合开发游戏。

4.风险上,Terminal-Bench 4.0为48%低于Kimi的67%,FrontierCode 2.0与Kimi同为26%但落后腾讯Hy4的31%;Humanity’s Last Exam落后Opus 5.5和GPT系列;腾讯Hy4单任务1.03美元但指数低8分。

总:事件应对上应区分长时程代理、专业文件生成与编程任务,优先在有案例的汽车、金融、工业流程试点,并关注安全测试结论。

总:工厂可关注Qwen4-Max在工业任务、自动化实验和3D设计中的生产设计需求与数字化机会。

1.生产设计上,Work360覆盖721个专业领域工业工作任务,模型得分52.69、单任务成本1.67美元;汽车领域SANF与Pacific Union单工件编程从约475分钟缩短到24分钟。

2.研发自动化上,与Deep Principle的Mira配合自主运行48小时,检索270万候选分子,控制自动化实验室完成四个设计合成测量循环,做出数据中心芯片冷却液;与深势科技PharmMaster优化24个候选分子,结合能力约为专利已知起始化合物的3.6倍。

3.数字化启示上,模型可调用90个MCP工具从多源拉数据,生成可校验Excel和PowerPoint,数字可追溯;与自研求解器MindOpt配合,数据标注算力分配成本平均降10.42%,分配CPU任务找最优解快14倍以上。

4.设计与制造上,可在Blender生成可复用于游戏和3D打印的三维模型;与Unity、Unreal Engine和Godot从单句自然语言描述搭建可玩游戏并自行运行调整。

总:工厂落地应先选有明确流程和可验证产出的环节,同时注意编程类基准并非全面领先。

总:服务商应看到模型效率提升带来的行业趋势,以及长任务交付、专业文件生成和自动化实验的解决方案。

1.趋势上,Qwen4-Max总参数2.4万亿、每次激活710亿、原生图像、100万token;每任务成本从4.43美元降至1.67美元,降62%,指数43分比前代高9分,价格约为前代三分之一。

2.新技术上,Qwen4架构用三层网络压缩前文、第四层只提取关键部分,加入四条并行连接和动态门控,3280亿参数N-gram嵌入层提供常见序列;长文本不再平均分配注意力,计算开销大降;Live Delivery支持初版交付和后续完善,插件开源。

3.客户痛点上,长任务结束时才暴露错误、长上下文成本高、多源数据整理难、工业流程效率低。解决方案包括用户中途介入改方向、调用90个MCP工具生成可追溯Excel和PowerPoint、汽车编程475分钟降至24分钟。

4.方案边界上,Terminal-Bench 4.0为48%低于Kimi的67%,FrontierCode 2.0为26%;但Agents Last Exam v1.1得分46.7最佳,物理实验可自主运行48小时并检索270万分子,需按场景组合。

总:服务商可优先围绕长时程代理、可校验文档生成和工业自动化包装交付。

总:平台商可关注专业代理模型对平台分发、工具生态和成本运营的需求,以及阿里最新平台做法。

1.平台需求上,用户需要低成本长上下文专业模型,Qwen4-Max单任务成本1.67美元,指数任务3.27美元,API输入1.50美元、输出4.50美元、缓存0.075美元;100万token和原生图像适合工业科研金融长任务。

2.最新做法上,通过QwenCloud上线,并在QwenWork、Qoder和QwenStudio提供入口;推出Live Delivery,先交付初版再迭代,用户可介入,插件开源;更小Qwen4计划以开源权重发布。

3.生态运营上,金融场景调用90个MCP工具,从多源拉数据并输出可校验Excel和PowerPoint;游戏开发与Unity、Unreal Engine和Godot配合;案例覆盖汽车、云基础设施、物理实验和药物研发。

4.风向规避上,Last Secure Code Bench功能完整且安全代码占比从22%升至31.8%,三个测试级别全部通过安全评级的代码占比从8%升至14.7%;化学、生物、放射与核风险经完整思维链审查,结论是未展现显著能力提升的真实创新。编程基准仍落后Kimi和腾讯Hy4,平台招商和任务准入需说明能力边界。

总:平台应以云API、工具入口和插件生态组织运营,同时用安全与基准边界管理预期。

总:研究者的产业新动向是超大规模稀疏模型通过架构创新实现成本下降,并进入代理、科研和工业工作流。

1.架构上,2.4万亿总参数、每次激活710亿,Qwen4混合三层压缩与第四层关键提取,四条并行连接和动态门控,3280亿参数N-gram嵌入;训练损失平稳下降,效率与成本改善有数据支撑。

2.基准问题上,Work360得分47.74升至52.69、成本4.43降至1.67美元;Artificial Analysis指数43分,比前代高9分,落后Kimi的46分;Terminal-Bench 4.0为48%低于Kimi的67%,FrontierCode 2.0为26%,但Agents Last Exam v1.1得分46.7最佳,HLE-Diamond升21分至45%。不同基准结论分化,需关注评测口径。

3.新问题与安全上,Last Secure Code Bench安全代码占比升到31.8%,三级全过14.7%;化学、生物、放射与核风险经内部评审完整思维链,认为未展现显著能力提升的真实创新。

4.商业模式上,按token收费、缓存折扣、Live Delivery代理工作流、MCP工具调用、QwenCloud及QwenWork等入口,小模型开源权重计划;案例覆盖汽车、金融、云基础设施、物理实验48小时和药物研发24个候选分子,显示专业工作流商业化路径。

总:研究上可继续关注混合注意力与稀疏设计、长时程代理评测、安全审查和开源权重对生态的影响。

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

Summary: Qwen4-Max’s most notable advances are stronger long-context and agent capabilities combined with significant cost reductions; ordinary users can use it directly.

1. The model has 2.4 trillion total parameters but activates only 71 billion per pass, natively supports image input, and has a 1 million-token context window, making it suitable for professional long-horizon tasks.

2. Cost per task falls from $4.43 to $1.67, a 62% reduction, while Work360 score rises from 47.74 to 52.69; API pricing is $1.50 per million input tokens, $4.50 for output, and $0.075 for cache hits.

3. In practice, it is available in QwenCloud, QwenWork, Qoder, and QwenStudio; Live Delivery ships a usable first version and improves it iteratively, with users able to intervene at any point and change direction, and the plug-in is open source.

4. Case studies show automotive single-part programming cut from about 475 minutes to 24 minutes; in finance, it can call 90 MCP tools to generate verifiable Excel and PowerPoint outputs with traceable numbers.

Bottom line: Its weaknesses also matter: Terminal-Bench 4.0 score is 48%, below Kimi K3.1 Pro’s 67%, and FrontierCode 2.0 is 26%, so selection should depend on the specific task.

Summary: For brands, the article presents Alibaba’s product positioning and competitive strategy of entering the market with a low-cost, high-performance professional model, which can inform pricing, channel, and use-case decisions.

1. On pricing, Qwen4-Max costs $3.27 per Intelligence Index task, about one-third of the previous generation’s $10.02 and roughly 11% below Kimi K3.1 Pro’s $3.68; API pricing is $1.50 input, $4.50 output, and $0.075 cache, with the price advantage coming from a 25% lower unit price, 70% cheaper cache, and lower token consumption.

2. On product development, the Qwen4 architecture combines three layers of compression with a fourth layer of key extraction, adds four parallel connections and dynamic gating, and includes a 328-billion-parameter N-gram embedding layer; long-context computation overhead falls and training is stable; native image support and 1 million tokens target industrial, research, and financial use cases.

3. On user behavior, the model is built around autonomous agents; Live Delivery ships a first version and then improves it, with users able to intervene and change direction, making it suitable for long tasks; automotive, financial, and game-development cases emphasize efficiency and verifiable deliverables.

4. On channels, it is already live on QwenCloud and available in QwenWork, Qoder, and QwenStudio; smaller Qwen4 models are planned for open-source weight releases, forming distribution through cloud APIs plus tool entry points.

Bottom line: In brand competition, it benchmarks against Opus 5.5, GPT-6, and Kimi K3.1 Pro. Its index score of 43 trails Kimi’s 46, but its Agents Last Exam v1.1 score of 46.7 is the best in the comparison group, so differentiation should be shaped by use case.

Summary: Sellers’ opportunity lies in lower costs for professional agent models and an open tool ecosystem; the risk is that some benchmarks trail competitors, so selection should be based on task and compliance.

1. On opportunity, per-task cost falls 62%, and Intelligence Index cost per task is $3.27, below Kimi’s $3.68; cache hits are 70% cheaper, API pricing is $1.50 input and $4.50 output, making it suitable for high-frequency long tasks.

2. On transferable patterns, Live Delivery ships a usable first version, iterates later, and allows user intervention, reducing the risk of errors surfacing only at the end of long tasks; in finance, the model calls 90 MCP tools to pull data and output traceable Excel and PowerPoint; the plug-in is open source.

3. On partnerships and support, it is already live on QwenCloud and available in QwenWork, Qoder, and QwenStudio; smaller Qwen4 models are planned for open-source weights; it works with Unity, Unreal Engine, and Godot for game development.

4. On risk, Terminal-Bench 4.0 is 48%, below Kimi’s 67%; FrontierCode 2.0 is 26%, tied with Kimi but behind Tencent Hy4’s 31%; Humanity’s Last Exam trails Opus 5.5 and the GPT series; Tencent Hy4 costs $1.03 per task but has an index 8 points lower.

Bottom line: In responding to the event, distinguish long-horizon agent tasks, professional document generation, and coding tasks; prioritize pilots in automotive, financial, and industrial workflows with existing cases, and monitor safety-test conclusions.

Summary: Factories can focus on Qwen4-Max’s production-design needs and digitalization opportunities in industrial tasks, automated experiments, and 3D design.

1. On production design, Work360 covers industrial work tasks across 721 professional domains; the model scores 52.69 at $1.67 per task; in automotive, SANF and Pacific Union cut single-part programming from about 475 minutes to 24 minutes.

2. On R&D automation, working with Deep Principle’s Mira, it runs autonomously for 48 hours, retrieves 2.7 million candidate molecules, and controls an automated lab to complete four design-synthesize-measure cycles, producing a data-center chip coolant; with DP Technology’s PharmMaster, it optimizes 24 candidate molecules, with binding affinity about 3.6 times that of known patented starting compounds.

3. On digitalization implications, the model can call 90 MCP tools to pull data from multiple sources and generate verifiable Excel and PowerPoint with traceable numbers; with the in-house solver MindOpt, it cuts data-labeling compute-allocation costs by an average of 10.42% and finds optimal solutions for CPU tasks more than 14 times faster.

4. On design and manufacturing, it can generate 3D models in Blender that can be reused for games and 3D printing; with Unity, Unreal Engine, and Godot, it builds playable games from a single natural-language description and runs and adjusts them autonomously.

Bottom line: Factory deployment should start with steps that have clear processes and verifiable outputs, while noting that coding benchmarks are not comprehensively ahead.

Summary: Service providers should note the industry trend driven by model efficiency gains, as well as solutions for long-task delivery, professional document generation, and automated experiments.

1. On the trend, Qwen4-Max has 2.4 trillion total parameters, activates 71 billion per pass, natively supports images, and offers a 1 million-token context; cost per task drops from $4.43 to $1.67, down 62%, while the index score of 43 is 9 points higher than the previous generation and price is about one-third of the predecessor.

2. On new technology, the Qwen4 architecture uses three network layers to compress prior context and a fourth layer to extract only key parts, adds four parallel connections and dynamic gating, and uses a 328-billion-parameter N-gram embedding layer to capture common sequences; long text no longer receives uniformly distributed attention, greatly reducing compute overhead; Live Delivery supports first-version delivery and later refinement, and the plug-in is open source.

3. On customer pain points, errors surface only at the end of long tasks, long-context costs are high, multi-source data is hard to organize, and industrial workflows are inefficient. Solutions include mid-task user intervention to change direction, calling 90 MCP tools to generate traceable Excel and PowerPoint, and cutting automotive programming from 475 minutes to 24 minutes.

4. On solution boundaries, Terminal-Bench 4.0 is 48%, below Kimi’s 67%, and FrontierCode 2.0 is 26%; however, Agents Last Exam v1.1 scores 46.7, the best, and physical experiments can run autonomously for 48 hours and retrieve 2.7 million molecules, so offerings should be combined by scenario.

Bottom line: Service providers can prioritize offerings around long-horizon agents, verifiable document generation, and industrial automation.

Summary: Marketplaces can focus on demand for distribution, tool ecosystems, and cost operations created by professional agent models, as well as Alibaba’s latest platform approach.

1. On platform demand, users need low-cost long-context professional models; Qwen4-Max costs $1.67 per task and $3.27 per Intelligence Index task, with API pricing of $1.50 input, $4.50 output, and $0.075 cache; 1 million tokens and native image support suit long industrial, research, and financial tasks.

2. On the latest approach, it is live on QwenCloud and offered through QwenWork, Qoder, and QwenStudio; Live Delivery ships a first version and then iterates, with user intervention possible, and the plug-in is open source; smaller Qwen4 models are planned for open-source weight releases.

3. On ecosystem operations, the financial scenario calls 90 MCP tools to pull data from multiple sources and output verifiable Excel and PowerPoint; game development works with Unity, Unreal Engine, and Godot; cases cover automotive, cloud infrastructure, physical experiments, and drug discovery.

4. On risk avoidance, Last Secure Code Bench shows functional completeness and a rise in secure-code share from 22% to 31.8%, while the share of code passing safety ratings across all three test levels rises from 8% to 14.7%; chemical, biological, radiological, and nuclear risks were reviewed through full chain-of-thought, with the conclusion that no genuine innovations showing significant capability gains were found. Coding benchmarks still trail Kimi and Tencent Hy4, so platform onboarding and task admission need to state capability boundaries.

Bottom line: Platforms should organize operations around cloud APIs, tool entry points, and plug-in ecosystems, while managing expectations with safety and benchmark boundaries.

Summary: The new industry development for researchers is that ultra-large sparse models are achieving cost reductions through architectural innovation and entering agent, research, and industrial workflows.

1. On architecture, the model has 2.4 trillion total parameters and activates 71 billion per pass; Qwen4 combines three layers of compression with a fourth layer of key extraction, four parallel connections and dynamic gating, and a 328-billion-parameter N-gram embedding; training loss declines steadily, and efficiency and cost improvements are supported by data.

2. On benchmarks, Work360 score rises from 47.74 to 52.69 and cost falls from $4.43 to $1.67; the Artificial Analysis index is 43, 9 points above the previous generation but behind Kimi’s 46; Terminal-Bench 4.0 is 48%, below Kimi’s 67%, and FrontierCode 2.0 is 26%, but Agents Last Exam v1.1 scores 46.7, the best, and HLE-Diamond rises 21 points to 45%. Conclusions differ across benchmarks, so evaluation methodology needs attention.

3. On new issues and safety, Last Secure Code Bench shows secure-code share rising to 31.8%, with 14.7% passing all three levels; chemical, biological, radiological, and nuclear risks were internally reviewed across full chain-of-thought, with the assessment that no genuine innovations showing significant capability gains were found.

4. On the business model, it uses token-based pricing, cache discounts, Live Delivery agent workflows, MCP tool calls, QwenCloud and QwenWork entry points, and plans for open-source weights for smaller models; cases cover automotive, finance, cloud infrastructure, 48-hour physical experiments, and 24 candidate molecules in drug discovery, showing a commercialization path for professional workflows.

Bottom line: Researchers can continue to watch hybrid attention and sparse design, long-horizon agent evaluation, safety review, and the impact of open-source weights on the ecosystem.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

阿里云发布新一代旗舰模型Qwen4-Max,总参数2.4万亿,每次推理仅激活710亿参数,原生支持图像输入,上下文窗口达到100万token。这款模型对标Anthropic的Opus 5.5、OpenAI的GPT-6以及月之暗面的Kimi K3.1 Pro,面向工业、科研与金融等专业工作场景,并以更低价格切入市场。

Qwen4-Max首次搭载重新设计的Qwen4架构。模型不再只依赖标准注意力层,而是混合两种机制。三层网络把前文压缩成浓缩状态,第四层只提取其中关键部分,长文本处理因此不再需要对上下文的每个部分平均分配注意力,计算开销大幅下降。架构还加入四条并行连接,配以动态门控稳定信息流动。一个含3280亿参数的N-gram嵌入层提供常见token序列作为现成构件。阿里公布的数据显示,2.4万亿参数配置的训练过程保持稳定,训练损失全程平稳下降,未出现突然跳变。

效率提升直接反映在成本上。在阿里自有的Work360基准中,该基准覆盖721个专业领域的工业工作任务,Qwen4-Max解决同类任务所用token比前代减少约四分之一到三分之一,单任务平均成本从4.43美元降至1.67美元,降幅62%,同期基准得分从47.74升至52.69。在代理基准Terminal-Bench 4.0上,token用量下降近一半。官方公告显示,Qwen4-Max在最高推理档位下于Artificial Analysis Intelligence Index拿到43分,比Qwen3.8-Max高出9分,落后中国模型榜首Kimi K3.1 Pro的46分。分项上,HLE-Diamond提升21分至45%,AA-Omniscience提升18分,AutomationBench-AA提升17分。

价格层面,Artificial Analysis测算Qwen4-Max每项Intelligence Index任务成本为3.27美元,约为Qwen3.8-Max的10.02美元的三分之一,比Kimi K3.1 Pro的3.68美元低约11%。节省来自三处,token单价较前代低25%,缓存命中便宜70%,同时模型消耗token更少,每项基准任务约需10.5万输出token,而前代为17.1万。腾讯的Hy4单任务成本低至1.03美元,但指数得分低8分。在Terminal-Bench 4.0上,Qwen4-Max得分48%,低于Kimi K3.1 Pro的67%。在Cognition的编程基准FrontierCode 2.0上,它与Kimi同为26%,落后腾讯Hy4的31%。通过自有API,阿里对每百万输入token收费1.50美元,输出token收费4.50美元,缓存命中0.075美元。

外部基准表现有高有低。在长时程代理任务Agents' Last Exam v1.1上,Qwen4-Max拿到46.7分,是对比组中最佳。它在Humanity's Last Exam上落后于Opus 5.5和GPT系列,在HealthBench Professional上落后于GPT系列但超过Opus 5.5。

Qwen4-Max围绕自主代理打造,并推出名为Live Delivery的功能。模型会先交付一个可用的初版,再在后续轮次中逐步完善,用户可以在任意节点介入并改变方向,避免长任务只在结束时才暴露错误而无法修正。配套的Qwen-Live-Delivery-Plugin已经开源。

阿里列举了若干落地案例。汽车领域的SANF与Pacific Union使用该模型后,单个工件的编程时间从约475分钟缩短到24分钟。金融场景中,模型调用90个MCP工具从多个来源拉取市场数据,输出可校验的Excel和PowerPoint文件,每个数字都可追溯回原始出处。阿里内部也用Qwen4-Max优化云基础设施,与自研求解器MindOpt配合后,数据标注的算力分配成本平均下降10.42%,在数据中心分配CPU任务时找到最优解的速度比对比方法快14倍以上。

Qwen4-Max还被用于物理实验。与创业公司Deep Principle的系统Mira配合后,模型自主运行48小时,设计、合成并测试了一款面向高性能数据中心芯片的冷却液,期间检索270万个候选分子,控制自动化实验室,完成四个完整的设计、合成与测量循环。在药物研发方向,模型与北京AI制药创业公司深势科技的平台PharmMaster协作,针对与阿尔茨海默病相关的炎症反应靶点,经过多轮文献检索、设计与计算,优化出24个候选分子。阿里公布,最终候选分子与靶蛋白的结合能力约为专利中已知起始化合物的3.6倍,深势科技专家认为该结果足以推进后续实验。

游戏开发是另一个展示场景。Qwen4-Max与Unity、Unreal Engine和Godot等成熟引擎配合,从单句自然语言描述出发,处理场景搭建、游戏逻辑、外部资源匹配以及材质、界面和视觉特效,直到游戏可玩。开发过程中模型会自行运行游戏,用视觉与交互反馈调整镜头、角色动画、受击反应和光照。阿里的示例项目是虚幻引擎冒险游戏Meow of Magic,讲述一只会魔法的猫对抗成波敌人的故事。Qwen4-Max还能在Blender中生成可复用于游戏和3D打印的三维模型。与之相反,Anthropic近期在大量非官方《光环》和《GTA:罪恶都市》克隆游戏出现后,禁止其编码助手Claude Code复刻现有游戏。

安全方面,在Last Secure Code Bench上,功能完整且安全的代码占比从22%升至31.8%,三个测试级别全部通过安全评级的代码占比从8%升至14.7%。阿里还对化学、生物、放射与核风险做了测试,内部评审人员检查了模型的完整思维链,结论是模型未展现出能为用户带来显著能力提升的真实创新能力。

Qwen4-Max已通过QwenCloud上线,并可在QwenWork、Qoder和QwenStudio中使用,更小的Qwen4模型计划后续以开源权重形式发布。在此之前,阿里于8月推出Qwen3.8-Max,同样为2.4万亿参数,但每token激活950亿参数,当时Artificial Analysis测得该模型每任务需64步而非14步,单任务成本是Qwen3.7-Max的两倍以上。8月下旬,阿里以Qwen3.8-Flash-Next预览Qwen4架构,其中已使用510亿参数的N-gram嵌入层。竞争对手方面,月之暗面10月推出Kimi K3.1 Pro,参数达2.8万亿,自称在Terminal-Bench 4.0上仅次于Claude Opus 5.5。DeepSeek在V4-Pro中采用混合注意力与稀疏注意力设计以降低长上下文成本,智谱AI的GLM-5.3则完全依靠额外后训练获得提升。

本文首发于 亿邦动力 官方网站

【版权提示】亿邦动力倡导尊重与保护知识产权。未经许可,任何人不得复制、转载、或以其他方式使用本网站的内容。如发现本站文章存在版权问题,烦请提供版权疑问、身份证明、版权证明、联系方式等发邮件至run@ebrun.com,我们将及时沟通与处理。

已点赞

+1

{{ praise_text }}

好文章,值得被更多人看到

FAQ回顾

Qwen4-Max是什么?有哪些核心技术特点?

Qwen4-Max是阿里云发布的新一代旗舰大模型,总参数2.4万亿,每次推理仅激活710亿参数,原生支持图像输入,上下文窗口达100万token。它首次搭载Qwen4架构,混合两种注意力机制、加入四条并行连接与动态门控,并配有3280亿参数的N-gram嵌入层,面向工业、科研与金融等专业工作场景。

Qwen4-Max的单任务成本是多少?成本为何能大幅下降?

Artificial Analysis测算Qwen4-Max每项Intelligence Index任务成本为3.27美元,约为Qwen3.8-Max的10.02美元的三分之一。在覆盖721个专业领域的Work360基准上,单任务平均成本从4.43美元降至1.67美元,降幅62%。节省来自三处:token单价较前代低25%、缓存命中便宜70%,且每项任务输出token从17.1万降至约10.5万。

Qwen4-Max在代理与编程类基准上表现如何?

Qwen4-Max在Terminal-Bench 4.0得分48%,低于Kimi K3.1 Pro的67%;在Cognition的FrontierCode 2.0上与Kimi同为26%,落后腾讯Hy4的31%。长时程代理任务Agents' Last Exam v1.1上拿到46.7分,为对比组最佳;Artificial Analysis Intelligence Index得43分,比Qwen3.8-Max高9分,但落后榜首Kimi K3.1 Pro的46分。

Qwen4-Max适合哪些行业场景?有哪些落地案例?

Qwen4-Max面向工业、科研与金融等专业场景。汽车领域SANF与Pacific Union将其用于工件编程,时间从约475分钟缩短到24分钟;金融场景中调用90个MCP工具拉取市场数据,输出数字可追溯的Excel与PowerPoint;与深势科技PharmMaster协作优化出24个阿尔茨海默病相关候选分子,结合能力约为专利已知起始化合物的3.6倍。

Qwen4-Max的API价格是多少?可以通过哪些渠道使用?

通过阿里自有API,Qwen4-Max每百万输入token收费1.50美元,输出token收费4.50美元,缓存命中0.075美元。模型已通过QwenCloud上线,并可在QwenWork、Qoder和QwenStudio中使用;更小的Qwen4模型计划后续以开源权重形式发布。

广告

下一篇

Akeneo AI代理白皮书:部署多代理如何形成运营模型

Akeneo白皮书警示:企业从首个AI代理到第十个代理时,挑战将从单一用例转向运营模型,需解决产品数据、治理与多代理协同问题,否则将放大碎片化数据风险。

零售商 AI 智慧零售 SaaS

2026-11-17 16:08:14