Models

Qwen4-Max 发布:2.4T 参数 MoE 架构,1M 上下文窗口

Why we picked it

阿里 Qwen 系列新旗舰,2.4T 参数加 1M 上下文,价格还降了。作者实测拿来批量分类上万封邮件和医疗收据。如果你的工作要塞长文档、大批量跑任务,这个便宜长窗口的组合挺实用。

Qwen4-Max 正式发布,采用 Mixture-of-Experts(MoE)架构,总参数量达 2.4T,上下文窗口扩展到 1M tokens,官方称价格比 Qwen3.8-Max 更低。MoE 架构意味着虽然总参数庞大,但每次推理只激活部分专家,从而在保持能力的同时控制推理成本。

Body

1M 上下文窗口配合更低定价,直接受益的是长文档处理类任务。作者举例:用该模型处理了 10,000+ 封邮件,识别其中的医疗收据并完成分类,这类任务 token 消耗极大,此前一直在等待一个「能力够强、价格够低」的模型组合。对需要批量处理长文本的企业(邮件、合同、票据归档)来说,1M 窗口意味着单次调用即可处理整批内容,不必做复杂的分块与拼接。

在超大参数 MoE 模型阵营里,Qwen4-Max 以低价和长上下文作为卖点,与主打性价比的路线一致。开发者若已有基于 Qwen 系列的流水线,长上下文任务可以优先试用这一代。

Original · Ate-a-Pi

Qwen4-Max is out!

• 2.4T-parameter model • 1M context window • Mixture-of-Experts (MoE) • Much cheaper than Qwen3.8-Max

I used the model to process 10,000+ emails, identify medical receipts, and classify them. It's a very token-heavy process, so I was waiting for something good + cheap.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Qwen4-Max, Qwen, MoE, reasoning models.

Source attribution: curated from Ate-a-Pi (@svpino).