据 Qwen 团队介绍,该混合专家(MoE)模型拥有 2.4 万亿参数,但每个 token 仅激活 710 亿参数。它原生支持图像处理,并提供一百万 token 的上下文窗口。
该架构还在模型层之间增加了四条并行连接,并配有动态门以稳定信息流。一个拥有 3280 亿参数的 N-gram 嵌入层提供常见的 token 序列作为现成的构建模块。据阿里巴巴称,训练这个 2.4 万亿参数的配置运行稳定。在整个训练过程中,训练损失稳步下降,没有出现突然的峰值。
阿里巴巴的内部评估通过知识、代码、问答和 STEM 领域的字节比特数来衡量模型质量。在测试的五次检查点中,Qwen4-Max-Base 得分最高。与 Qwen3.8-Max-Base 相比,其性能提升了 5% 到 13%,同时每个 token 激活的参数减少了 25%。
效率提升体现在运营成本上。Work360 是阿里巴巴针对 721 个专门领域工业工作任务的自有基准测试。在该测试中,Qwen4-Max 解决相同任务所需的 token 数量比其前身减少了约四分之一到三分之一。在代理基准测试 Terminal-Bench 4.0 上,token 使用量几乎减半。
据阿里巴巴称,Work360 任务的平均成本下降了 62%,从 4.43 美元降至 1.67 美元。与此同时,基准测试分数从 47.74 分上升到 52.69 分。
外部基准测试的结果喜忧参半。在 Agents' Last Exam v1.1 等长周期代理任务中,Qwen4-Max 得分为 46.7 分,在对比组中表现最佳。但在 Humanity's Last Exam 中,它落后于 Opus 5.5 和 GPT 系列模型。在 HealthBench Professional 中,它落后于 GPT 系列模型,但略胜 Opus 5.5。
阿里巴巴表示,Qwen4-Max 在代理和工作任务方面领先,而 Opus 5.5 和 GPT 系列模型在 Humanity's Last Exam 中表现更好。| 图片来源:阿里巴巴
在其最高推理设置下,Qwen4-Max 在当前的 Artificial Analysis Intelligence Index 中得分为 43 分。这比 Qwen3.8-Max 高出 9 分,但落后于排名第一的中国模型 Kimi K3.1 Pro(46 分)。与前身相比,最大的提升来自 HLE-Diamond(上升 21 点至 45%)、AA-Omniscience(上升 18 分)和 AutomationBench-AA(上升 17 分)。
Qwen4-Max 在 Artificial Analysis Intelligence Index 中增加了 9 分,且每项任务的成本仅为前身的约三分之一。| 图片来源:Artificial Analysis
Artificial Analysis 将 Qwen4-Max 的成本定为每项 Intelligence Index 任务 3.27 美元。这大约是 Qwen3.8-Max 10.02 美元的三分之一,也略低于 Kimi K3.1 Pro 的 3.68 美元(低约 11%)。节省主要来自三个方面。Token 价格比前身低 25%,缓存命中成本低 70%,且模型消耗的 token 更少。Qwen4-Max 每项基准任务需要约 105,000 个输出 token,而 Qwen3.8-Max 需要 171,000 个。
与 AI 行业的许多做法一样,阿里巴巴围绕能够独立处理长时间任务的代理构建了此次发布。如果用户只能在最后看到结果,他们几乎没有办法纠正错误。阿里巴巴的解决方案是“Live Delivery”(实时交付)。模型早期交付一个可用的初版,并在后续迭代中进行完善。用户可以随时介入并改变方向。阿里巴巴已开源了配套的 Qwen-Live-Delivery-Plugin。
阿里巴巴引用了几个合作伙伴部署案例。据报道,汽车合作伙伴 SANF 和 Pacific Union 使用 Qwen4-Max 将单个工件的手动编程时间从约 475 分钟缩短至 24 分钟。在金融领域,该模型调用 90 个 MCP 工具从多个来源提取市场数据。它以可验证的 Excel 和 PowerPoint 文件形式交付结果,且每个数字都能追溯回其原始来源。
阿里巴巴还内部使用该模型以更高效地运行其自身的云基础设施。据该公司称,Qwen4-Max 与内部求解器 MindOpt 合作,将数据标注计算作业分配的平均成本降低了 10.42%。在数据中心分发 CPU 任务时,它找到最优解的速度比对比方法快 14 倍以上。
阿里巴巴还展示了 Qwen4 系列在物理实验室中的应用。据阿里巴巴称,与初创公司 Deep Principle 的系统 Mira 配合使用时,该模型自主运行了 48 小时。在此期间,它设计、合成并测试了一种用于高性能数据中心芯片的新型冷却液。它搜索了 270 万个分子候选者,控制自动化实验室,并完成了四轮完整的设计、合成和测量循环。
Qwen4-Max 还与位于北京的 AI 药物发现初创公司 DP Technology 的平台 PharmMaster 合作。目标是针对与阿尔茨海默病相关的炎症反应的化合物。经过几轮文献综述、设计和计算,该模型优化了 24 个候选分子。据阿里巴巴称,最终候选者与已知专利起始化合物相比,与目标蛋白的结合能力提高了约 3.6 倍。DP Technology 的专家认为这一结果值得进一步开展实验室测试。
阿里巴巴还展示了 Qwen4-Max 与 Unity、Unreal Engine 和 Godot 等成熟游戏引擎的合作。根据单个自然语言描述,该模型应能处理整个开发过程。这包括场景设置、游戏逻辑和匹配外部资产,以及材质、UI 和视觉效果,直到游戏可玩为止。
在开发过程中,模型自身运行游戏,并利用视觉和交互反馈来调整摄像机运镜、角色动画、受击反应和光照。阿里巴巴的示例项目是“Meow of Magic”,一款关于猫巫师的 Unreal Engine 冒险游戏,玩家在其中施放法术并与一波波敌人战斗。据阿里巴巴称,Qwen4-Max 还能在 Blender 中构建可用于游戏和 3D 打印的 3D 模型。
据阿里巴巴称,在安全基准测试 Last Secure Code Bench 上,既功能正常又安全的代码比例从 22% 上升至 31.8%。在三个测试级别均被评为安全的代码比例几乎翻了一番,从 8% 上升至 14.7%。阿里巴巴还测试了 CBRN 风险(化学、生物、放射性和核武器)。该公司表示,内部审查员审查了模型的完整思维链,未发现任何能显著增强用户能力的创新潜力。
查看原文
Qwen4-Max is Alibaba's answer to Opus 5.5, GPT-6, and Kimi K3.1 Pro. The new flagship pairs a more efficient architecture with much lower prices and targets professional work in industry, science, and finance.
The mixture-of-experts model has 2.4 trillion parameters but activates only 71 billion per token, according to the Qwen team. It processes images natively and offers a one-million-token context window.
A hybrid architecture makes long contexts cheaper to process
Qwen4-Max debuts the revamped architecture of the Qwen4 series. Instead of relying only on standard attention layers, Alibaba combines two methods. Three layers compress the prior context into a condensed state, and a fourth layer then picks out only the parts that matter. That sharply cuts compute costs on long inputs, since the model no longer has to give every part of the context equal attention.

Qwen4-Max compresses prior text across three layers before a fourth layer takes a closer look at only the important sections. | Image: Alibaba
The architecture also adds four parallel connections between model layers, with dynamic gates meant to stabilize the flow of information. An N-gram embedding layer with 328 billion parameters supplies common token sequences as ready-made building blocks. According to Alibaba, training the 2.4-trillion-parameter configuration ran stably. Training loss fell steadily throughout the run with no sudden spikes.

Training loss declined steadily throughout the run with no loss spikes, according to Alibaba. | Image: Alibaba
Alibaba's internal evaluations measure model quality in bits per byte across knowledge, code, question answering, and STEM. Qwen4-Max-Base scored best among the five checkpoints tested. It improved 5 to 13 percent over Qwen3.8-Max-Base while activating 25 percent fewer parameters per token.

Qwen4-Max-Base posts the lowest, and therefore best, values in all four categories. | Image: Alibaba
Lower token use cuts the cost of each task by more than half
The efficiency gains show up in operating costs. Work360 is Alibaba's own benchmark for industrial work tasks across 721 specialized fields. On it, Qwen4-Max uses roughly a quarter to a third fewer tokens than its predecessor to solve the same tasks. On the agent benchmark Terminal-Bench 4.0, token use drops by nearly half.
According to Alibaba, the average cost per Work360 task falls 62 percent, from $4.43 to $1.67. The benchmark score climbs from 47.74 to 52.69 at the same time.

Qwen4-Max beats its predecessor on both benchmarks while using far fewer tokens. | Image: Alibaba
Results on external benchmarks are mixed. On long-horizon agent tasks like Agents' Last Exam v1.1, Qwen4-Max scores 46.7 points, the best result in the comparison group. It trails Opus 5.5 and the GPT models on Humanity's Last Exam. On HealthBench Professional, it falls behind the GPT models but edges past Opus 5.5.

Alibaba says Qwen4-Max leads on agent and work tasks, while Opus 5.5 and the GPT models come out ahead on Humanity's Last Exam. | Image: Alibaba
Qwen4-Max ranks second among Chinese models at a third of its predecessor's cost
At its highest reasoning setting, Qwen4-Max scores 43 points on the current Artificial Analysis Intelligence Index. That's nine more than Qwen3.8-Max and three behind Kimi K3.1 Pro, the top Chinese model at 46 points. The biggest gains over the predecessor come on HLE-Diamond (up 21 points to 45 percent), AA-Omniscience (up 18), and AutomationBench-AA (up 17).

Qwen4-Max gains nine points on the Artificial Analysis Intelligence Index and costs only about a third as much per task as its predecessor. | Image: Artificial Analysis
Artificial Analysis puts the cost of Qwen4-Max at $3.27 per Intelligence Index task. That's about a third of Qwen3.8-Max's $10.02 and roughly 11 percent less than Kimi K3.1 Pro's $3.68. The savings come from three places. Token prices are 25 percent lower than the predecessor's, cache hits cost 70 percent less, and the model burns through fewer tokens. Qwen4-Max needs about 105,000 output tokens per benchmark task, compared with 171,000 for Qwen3.8-Max.
Tencent's Hy4 is much cheaper at $1.03 per task but scores eight points lower on the index. On Terminal-Bench 4.0, Qwen4-Max scores 48 percent, behind Kimi K3.1 Pro at 67 percent. On Cognition's coding benchmark FrontierCode 2.0, it ties Kimi at 26 percent and trails Tencent Hy4 at 31 percent.
Through its own API, Alibaba charges $1.50 per million input tokens and $4.50 per million output tokens. Cache hits cost $0.075.
Live Delivery hands control of long-running agents back to users
Like much of the AI industry, Alibaba builds this release around agents that handle long tasks on their own. If users only see results at the end, they have almost no way to fix mistakes. Alibaba's answer is "Live Delivery." The model delivers a usable first version early and builds it out in later passes. Users can step in and change course at any point. Alibaba has open-sourced the accompanying Qwen-Live-Delivery-Plugin.
Alibaba cites several partner deployments. Automotive partners SANF and Pacific Union reportedly used Qwen4-Max to cut manual programming time for a single workpiece from about 475 minutes to 24. In finance, the model taps 90 MCP tools to pull market data from multiple sources. It delivers the results as verifiable Excel and PowerPoint files, and every number traces back to its original source.
Alibaba also uses the model internally to run its own cloud infrastructure more efficiently. According to the company, Qwen4-Max worked with the in-house solver MindOpt to cut the cost of allocating compute jobs for data annotation by an average of 10.42 percent. When distributing CPU tasks in data centers, it found optimal solutions more than 14 times faster than the comparison method.
Alibaba's models now run lab experiments on their own
Alibaba also shows the Qwen4 generation at work in physical labs. Paired with Mira, a system from startup Deep Principle, the model ran autonomously for 48 hours, according to Alibaba. In that time, it designed, synthesized, and tested a new coolant for high-performance data center chips. It searched 2.7 million molecular candidates, controlled an automated lab, and completed four full cycles of design, synthesis, and measurement.
Qwen4-Max also worked with PharmMaster, a platform from Beijing-based AI drug discovery startup DP Technology. The target was a compound aimed at an inflammatory response linked to Alzheimer's disease. Over several rounds of literature review, design, and computation, the model optimized 24 candidate molecules. According to Alibaba, the final candidate binds to the target protein about 3.6 times better than the known starting compound from a patent. DP Technology's experts consider the result promising enough to pursue further lab tests.
Qwen4-Max aims to build playable games from a single prompt
Alibaba also shows Qwen4-Max working with established game engines like Unity, Unreal Engine, and Godot. From a single natural-language description, the model is supposed to handle the entire development process. That covers scene setup, game logic, and matching external assets, along with materials, UI, and visual effects, until the game is playable.

Qwen4-Max built a wizard adventure in Unreal Engine in which players, as a cat, battle waves of enemies and bosses. | Image: Alibaba
During development, the model runs the game itself and uses visual and interactive feedback to tweak camera work, character animations, hit reactions, and lighting. Alibaba's sample project is "Meow of Magic," an Unreal Engine adventure about a cat wizard in which players cast spells and fight waves of enemies. According to Alibaba, Qwen4-Max also builds 3D models in Blender that can be reused for games and 3D printing.
Anthropic recently moved its coding assistant Claude Code in the opposite direction. After a wave of unofficial Halo and GTA: Vice City clones, it barred the tool from recreating existing games.
Alibaba says the model writes safer code and offers no real CBRN uplift
On the safety benchmark Last Secure Code Bench, the share of code that is both functional and secure rose from 22 to 31.8 percent, according to Alibaba. The share rated secure at all three testing levels nearly doubled, from 8 to 14.7 percent. Alibaba also tested for CBRN risks (chemical, biological, radiological, and nuclear). Internal reviewers went through the model's full chains of thought, the company says. They found no real capacity for innovation that would give a user a significant boost in capability.
Qwen4-Max is available through QwenCloud and in QwenWork, Qoder, and QwenStudio. Alibaba plans to release smaller Qwen4 models as open source later.
In August, Alibaba released Qwen3.8-Max, its first Max model built for autonomous tasks that run for days. It also has 2.4 trillion parameters, but 95 billion are active per token. At the time, Artificial Analysis found that the model needed 64 steps per task instead of 14 and cost more than twice as much per task as Qwen3.7-Max.
In late August, Alibaba previewed the Qwen4 architecture with Qwen3.8-Flash-Next, which already used an N-gram embedding layer with 51 billion parameters. The company had released the Qwen3.8-Max weights in mid-August. For Qwen4, only the smaller models will get open weights for now.
Moonshot AI followed in October with Kimi K3.1 Pro. At 2.8 trillion parameters, it's even larger, and Moonshot says it sits just behind Claude Opus 5.5 on Terminal-Bench 4.0. Like Alibaba, DeepSeek uses a hybrid attention design with sparse attention in V4-Pro to process long contexts more cheaply. Zhipu AI took a different route and got GLM-5.3's gains entirely from extra post-training.