阿里发布Qwen4-Max:2.4T参数MoE,API降价62%
Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M Context
Qwen4旗舰发布且伴随显著降价,对Agent工作流有直接吸引力,建议开发者关注其成本优势与长上下文能力。
Alibaba’s Qwen team has released Qwen4-Max, the first model of its Qwen4 generation. Qwen4-Max is a proprietary sparse MoE model with a 2.4T-parameter backbone and 71B parameters active per token. It accepts text, images and video, and supports a 1M-token context window. Alibaba also cut API prices and reports a 62% drop in average cost per task.
阿里巴巴 Qwen 团队发布了 Qwen4-Max,这是其 Qwen4 系列的首个模型。Qwen4-Max 是一款专有的稀疏 MoE(混合专家)模型,拥有 2.4T 参数的主干网络,每个 token 激活 71B 参数。它支持文本、图像和视频输入,并支持 1M token 的上下文窗口。阿里巴巴还降低了 API 价格,并表示平均任务成本下降了 62%。
TL;DR
TL;DR(太长不看版)
- Size: 2.4T backbone parameters, 71B active per token, plus 328B N-gram Embedding parameters. 1M context, 393K max output tokens.
- Runs on: Hosted only: QwenCloud API, Token Plan, Qwen Studio, QwenWork and Qoder. No open weights, so no self-hosting or quantized builds.
- Performance: A large jump over Qwen3.8-Max. Competitive with frontier models on agentic work, behind on hard coding and vision.
- Best: HLAB-Vals-1.1 (legal agent work): 21.3, versus 11.7 for Kimi K3.1, the next-best model.
- 规模:2.4T 主干参数,每个 token 激活 71B,外加 328B N-gram 嵌入参数。1M 上下文窗口,最大输出 393K token。
- 运行方式:仅限托管:QwenCloud API、Token Plan、Qwen Studio、QwenWork 和 Qoder。未开源权重,因此无法自托管或构建量化版本。
- 性能:相比 Qwen3.8-Max 有大幅提升。在代理工作方面与前沿模型竞争,但在硬编码和视觉任务上稍逊一筹。
- 最佳表现:HLAB-Vals-1.1(法律代理工作):21.3,而次优模型 Kimi K3.1 为 11.7。
What is Qwen4-Max?
什么是 Qwen4-Max?
Qwen4-Max is a multimodal MoE language model from Alibaba, built for long-horizon professional work. A Token Plan-only preview opened this month only.
Qwen4-Max 是阿里巴巴推出的一款多模态 MoE 语言模型,专为长周期专业工作设计。本月仅通过 Token Plan 开放预览。
How does the Qwen4-Max architecture work?
Qwen4-Max 的架构是如何工作的?
Qwen4 upgrades the sparse MoE design in 3 places: attention, residual connections and embeddings.
Qwen4 在三个地方升级了稀疏 MoE 设计:注意力机制、残差连接和嵌入层。
- GDN + Qwen Sparse Attention (QSA): Each QSA layer pairs with 3 Gated DeltaNet layers. These compress history into fixed-size states. QSA then uses a lightweight indexer to pick relevant context at micro-block granularity. This lowers attention and indexing cost on long sequences.
- Gated Residual (GR): 4 parallel residual branches use dynamic read and write gates. The aim is stronger information flow and better training stability.
- N-gram Embedding: 328B extra parameters provide lookups of local token patterns at layer 2. This adds capacity with little extra compute.
- GDN + Qwen 稀疏注意力(QSA):每个 QSA 层与 3 个门控 DeltaNet 层配对。这些层将历史信息压缩为固定大小的状态。随后,QSA 使用轻量级索引器以微块粒度选择相关上下文。这降低了长序列上的注意力和索引成本。
- 门控残差(GR):4 个并行残差分支使用动态读写门控。旨在增强信息流动并提高训练稳定性。
- N-gram 嵌入:额外的 328B 参数在第 2 层提供局部 token 模式的查找。这在几乎不增加额外计算量的情况下增加了容量。
Alibaba reports no loss spikes across the 2.4T-scale pre-training run. On in-house tests, Qwen4-Max-Base shows 5 to 13% lower bits per byte than Qwen3.8-Max-Base. It does this while activating 25% fewer parameters.
阿里巴巴报告称,在 2.4T 规模的预训练过程中没有出现损失尖峰。在内部测试中,Qwen4-Max-Base 的每字节比特数比 Qwen3.8-Max-Base 低 5% 到 13%。这是在激活参数减少 25% 的情况下实现的。
How does Qwen4-Max perform on benchmarks?
Qwen4-Max 在基准测试中的表现如何?
Alibaba’s comparison chart covers 16 benchmarks. Rivals are Qwen3.8-Max-0902, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1. Several scores are cited from third-party leaderboards. None were independently reproduced at launch.
阿里巴巴的比较图表涵盖了 16 项基准测试。竞争对手包括 Qwen3.8-Max-0902、Claude Opus 5.5、GPT-6.1 Sol、GPT-6 Astra 和 Kimi K3.1。部分分数引用自第三方排行榜。发布时均未进行独立复现。
Qwen4-Max ranks first on 4 of 16 benchmarks:
Qwen4-Max 在 16 项基准测试中的 4 项排名第一:
- HLAB-Vals-1.1 (legal agents): 21.3, versus 11.7 for Kimi K3.1 and 9.2 for Opus 5.5.
- Agents’ Last Exam v1.1: 46.7, versus 43.4 for Opus 5.5 and Kimi K3.1.
- BioMysteryBench: 85.0, versus 84.4 for GPT-6.1 Sol.
- DeepSWE v1.1: 75.7, versus 75.5 for Kimi K3.1.
- HLAB-Vals-1.1(法律代理):21.3,而 Kimi K3.1 为 11.7,Opus 5.5 为 9.2。
- Agents’ Last Exam v1.1:46.7,而 Opus 5.5 和 Kimi K3.1 均为 43.4。
- BioMysteryBench:85.0,而 GPT-6.1 Sol 为 84.4。
- DeepSWE v1.1:75.7,而 Kimi K3.1 为 75.5。
Claude Opus 5.5 leads 7 of the 16 charts. Qwen4-Max trails most clearly on coding and vision:
Claude Opus 5.5 在 16 项指标中领先 7 项。Qwen4-Max 在编程和视觉方面明显落后:
- ProgramBench: 31.5 versus 65.0 for Opus 5.5.
- Terminal-Bench 4.0: 48.5 versus 67.2 for Kimi K3.1.
- Chartography: 54.2 versus 71.1 for GPT-6 Astra.
- Humanity’s Last Exam: 51.7 versus 59.8 for Opus 5.5.
- ProgramBench:Opus 5.5 得分为 65.0,而 Qwen4-Max 为 31.5。
- Terminal-Bench 4.0:Kimi K3.1 得分为 67.2,而 Qwen4-Max 为 48.5。
- Chartography:GPT-6 Astra 得分为 71.1,而 Qwen4-Max 为 54.2。
- Humanity’s Last Exam:Opus 5.5 得分为 59.8,而 Qwen4-Max 为 51.7。
On Alibaba’s own Work360 benchmark, Qwen4-Max scores 52.69, up from 47.74 for Qwen3.8-Max. Work360 spans 26 top-level domains and 721 specialized fields.
在阿里巴巴自家的 Work360 基准测试中,Qwen4-Max 得分为 52.69,高于 Qwen3.8-Max 的 47.74。Work360 涵盖 26 个顶级领域和 721 个专业领域。
Why is Qwen4-Max cheaper to run?
为什么 Qwen4-Max 的运行成本更低?
Two things stack: fewer tokens and lower prices. On Work360, token use drops about 24% at Xhigh effort and 35% at High effort. On Terminal-Bench 4.0, the drops are about 48% and 49%. Average Work360 cost per task falls from $4.43 to $1.67, a 62% cut.
两个因素叠加:Token 用量更少且价格更低。在 Work360 上,在高努力程度下 Token 用量减少约 24%,在较高努力程度下减少 35%。在 Terminal-Bench 4.0 上,降幅分别约为 48% 和 49%。每个任务的平均 Work360 成本从 4.43 美元降至 1.67 美元,降幅达 62%。
List prices fell too. Input is $1.50 per million tokens, down from $2 for Qwen3.8-Max. Output is $4.50, down from $6. Implicit cache reads drop 70%, from $0.25 to $0.075.
列表价格也随之下降。输入价格为每百万 Token 1.50 美元,低于 Qwen3.8-Max 的 2 美元。输出价格为 4.50 美元,低于 6 美元。隐式缓存读取量下降 70%,从 0.25 美元降至 0.075 美元。
Qwen4-Max vs competitors
Qwen4-Max 与竞争对手对比
| Metrics | Qwen4-Max | Qwen3.8-Max | Claude Opus 5.5 | GPT-6.1 Sol | Kimi K3.1 Pro |
|---|---|---|---|---|---|
| Total params | 2.4T backbone + 328B N-gram emb. | 2.4T | Not disclosed | Not disclosed | 2.8T |
| Active params | 71B | Not disclosed (blog implies about 95B) | Not disclosed | Not disclosed | 104B |
| Context | 1M | 1M | 1M | 1,050,000 | 1,048,576 |
| Max output | 393K | 131K | 128K | 128K | Not disclosed |
| Input modalities | Text, image, video | Not disclosed | Text, image | Text, image | Text, image |
| Weights | Proprietary, API only | Not disclosed | Proprietary | Proprietary | Open (Kimi K3.1 Pro License) |
| Input / output price (per 1M) | $1.50 / $4.50 | $2 / $6 | $4 / $20 | $2 / $10 | $2 / $10 |
| Cached input (per 1M) | $0.075 | $0.25 | $0.20 | $0.10 | $0.20 |
| DeepSWE v1.1* | 75.7 | 69.3 | 74.2 | 75.2 | 75.5 |
| Terminal-Bench 4.0* | 48.5 | 38.9 | 59.6 | 56.1 | 67.2 |
| Humanity’s Last Exam* | 51.7 | 43.9 | 59.8 | 54.0 | 51.1 |
| ProgramBench* | 31.5 | 28.0 | 65.0 | 52.5 | 25.0 |
| HLAB-Vals-1.1* | 21.3 | 15.4 | 9.2 | 6.3 | 11.7 |
| 指标 | Qwen4-Max | Qwen3.8-Max | Claude Opus 5.5 | GPT-6.1 Sol | Kimi K3.1 Pro |
|---|---|---|---|---|---|
| 总参数量 | 2.4T 主干 + 328B N-gram 嵌入 | 2.4T | 未披露 | 未披露 | 2.8T |
| 激活参数量 | 71B | 未披露(博客暗示约 95B) | 未披露 | 未披露 | 104B |
| 上下文窗口 | 1M | 1M | 1M | 1,050,000 | 1,048,576 |
| 最大输出长度 | 393K | 131K | 128K | 128K | 未披露 |
| 输入模态 | 文本、图像、视频 | 未披露 | 文本、图像 | 文本、图像 | 文本、图像 |
| 权重类型 | 专有,仅限 API | 未披露 | 专有 | 专有 | 开源(Kimi K3.1 Pro 许可证) |
| 输入/输出价格(每百万) | $1.50 / $4.50 | $2 / $6 | $4 / $20 | $2 / $10 | $2 / $10 |
| 缓存输入(每百万) | $0.075 | $0.25 | $0.20 | $0.10 | $0.20 |
| DeepSWE v1.1* | 75.7 | 69.3 | 74.2 | 75.2 | 75.5 |
| Terminal-Bench 4.0* | 48.5 | 38.9 | 59.6 | 56.1 | 67.2 |
| Humanity’s Last Exam* | 51.7 | 43.9 | 59.8 | 54.0 | 51.1 |
| ProgramBench* | 31.5 | 28.0 | 65.0 | 52.5 | 25.0 |
| HLAB-Vals-1.1* | 21.3 | 15.4 | 9.2 | 6.3 | 11.7 |
*Scores from Alibaba’s Qwen4-Max chart. Qwen3.8 column is Qwen3.8-Max-0902. The chart labels Kimi as “Kimi K3.1”. Moonshot’s own card reports Terminal-Bench 4.0 of 65.7 for Kimi K3.1 Pro and 66.4 for Opus 5.5.
*分数来自阿里巴巴的 Qwen4-Max 图表。Qwen3.8 列对应的是 Qwen3.8-Max-0902。图表中将 Kimi 标记为“Kimi K3.1”。月之暗面(Moonshot)自己的卡片报告称,Kimi K3.1 Pro 在 Terminal-Bench 4.0 上的得分为 65.7,Opus 5.5 为 66.4。
How can developers use Qwen4-Max?
开发者如何使用 Qwen4-Max?
The model ID is qwen4-max on QwenCloud. The API works with OpenAI-compatible and Anthropic-compatible clients. Supported features include function calling, structured outputs, context caching, batch processing, web search and fine-tuning. The default rate limit is 5,000,000 tokens per minute. Thinking mode is toggled with enable_thinking.
模型 ID 为 QwenCloud 上的 qwen4-max。API 兼容 OpenAI 和 Anthropic 客户端。支持的功能包括函数调用、结构化输出、上下文缓存、批处理、网络搜索和微调。默认速率限制为每分钟 5,000,000 个 Token。思考模式通过 enable_thinking 进行切换。
Subscription access runs through the Token Plan.
订阅访问通过 Token Plan 实现。
Key Takeaways
关键要点
- 2.4T MoE backbone, only 71B active per token.
- New QSA attention, Gated Residuals and 328B N-gram Embeddings.
- First on 4 of 16 charted benchmarks, led by HLAB-Vals-1.1 at 21.3.
- $1.50/$4.50 per million tokens, cached input down 70%.
- Weak spots: ProgramBench (31.5) and Terminal-Bench 4.0 (48.5).
- 2.4T MoE 主干网络,每个 token 仅激活 71B 参数。
- 全新 QSA 注意力机制、门控残差连接以及 328B N-gram 嵌入。
- 在列出的 16 项基准测试中首次取得 4 项第一,以 HLAB-Vals-1.1(得分 21.3)领先。
- 每百万 token 费用为 $1.50/$4.50,缓存输入降低 70%。
- 薄弱环节:ProgramBench(31.5)和 Terminal-Bench 4.0(48.5)。
Check out the official Qwen4-Max blog, the QwenCloud model page and the Token Plan pricing page. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.
请参阅官方 Qwen4-Max 博客、QwenCloud 模型页面以及 Token Plan 定价页面。您可以通过我们的 GPULeaderboard.com Chrome 扩展程序在此查看部署选项和规格。
Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
此外,欢迎在 Twitter 上关注我们,别忘了加入拥有 15 万+成员的 ML SubReddit 并订阅我们的新闻通讯。等等!你也在 Telegram 上吗?现在你也可以加入我们的 Telegram 群组。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力