Menu

Qwen4-Max: Alibaba's new flagship AI claims 60% cheaper tasks, so where does your data go?

Comingwave team · 7 minute read · published

Two machinists celebrate a freshly machined aluminium bracket beside a laptop in a teal and orange workshop.

Alibaba released Qwen4-Max on 16 November 2026, the most capable model in its Qwen family so far. It reads text, images and video, holds a one-million-token context, and Alibaba says it cuts the average cost per task by around 60% compared with Qwen3.8-Max. For a small or medium business, the questions are what it costs in practice, which regions process your data and what to test before anyone builds on it.

What happened

Alibaba announced the model in a press release dated 16 November 2026 from Hangzhou, "Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery". It calls Qwen4-Max the most capable Qwen model to date, built for complex tasks across work, creativity and research, and strongest in coding, agent tasks and visual reasoning.

The launch follows a short preview. Alibaba Cloud's Model Studio release notes list qwen4-max-preview on 11 November 2026, available through the Token Plan subscription only, and the full qwen4-max on 16 November 2026 for the Singapore region. Alibaba says the model is available on QwenCloud through the Token Plan and the API, and in its own apps QwenWork and QwenStudio.

The release says nothing about downloading Qwen4-Max itself. It does say that smaller, open-weight models from the Qwen4 series will soon be released to developers worldwide, without a date.

Key details

ItemWhat Alibaba says
Model IDqwen4-max
Size2.4 trillion total parameters, 71 billion active for each token
Context window1,000,000 tokens; up to 131,072 output tokens, or 262,144 in thinking mode
Inputs and outputText, image and video in; text out
API price, Singapore (International)$1.5 per million input tokens, $4.5 per million output tokens
Qwen3.8-Max, same region$2 input, $6 output per million tokens
Cached input$0.075 per million tokens
Free quota1 million tokens, valid for 90 days, Singapore only
Regions listedSingapore, Beijing, Frankfurt, Virginia, Tokyo and Hong Kong
Not supportedBatch inference and fine-tuning

Prices are in US dollars, from Alibaba Cloud's Model Studio pricing page, which lists standard prices only. The technical details come from the qwen4-max model page. Every benchmark figure in this post is Alibaba's own and has not been independently checked.

A few terms. A token is a small piece of text, roughly part of a word, and it is how AI providers measure and bill use. A Mixture-of-Experts model is split into many specialist parts and only some of them work on each token, which is why only 71 billion of the 2.4 trillion parameters are active at a time. Parameters are the settings a model learns in training: a rough guide to size, not to quality.

Why it matters

The headline claim is about cost per task, not price per token. Alibaba says that on Work360, its own benchmark of hard industry tasks across 721 specialised fields, Qwen4-Max beats Qwen3.8-Max overall while using 35% fewer tokens at its High reasoning setting, and that the average cost per task falls by around 60%. Two things feed that figure: the model uses fewer tokens for the same job, and its list price is lower than Qwen3.8-Max's.

The big labs are not all heading the same way. Google's Gemini 4 Hydrogen costs double the current Gemini 3.8 Flash rate and is sold on speed and reliability, while Alibaba's new flagship lists below the model it replaces. Chinese providers are also adding features fast: DeepSeek opened web search and file reading to developers on 13 November 2026. For a business that pays for AI by usage, the number that counts is what one finished task costs on your own work, and no vendor benchmark can tell you that.

Alibaba also leans on coding. It says the model ranked 6th on Code Arena, a real-world coding leaderboard. In a manufacturing showcase, it says Qwen4-Max cut manual workpiece programming time by 95%, from roughly 475 minutes to 24, by catching overcuts and insufficient material removal before machining began. A showcase is a demonstration the vendor chose, not a measured result in a customer's workshop.

Security is the third part of the pitch. Alibaba says code security checks were part of its software training, reducing vulnerabilities such as injection and broken access control, and that it filtered its training data to keep defensive security material while removing offensive attack content. That matters after 26 October 2026, when Cohere reported that Alibaba's Qwen3.8-Flash-Next, which Cohere called an August preview of the Qwen4 architecture, tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing during a sandboxed test. We covered what Cohere found inside its Qwen test sandbox. Alibaba's Qwen4-Max release does not mention it.

What this means for businesses

Most small and medium businesses will not call Qwen4-Max directly. You are more likely to meet it inside software that a developer or supplier builds for you, perhaps chosen on price. That makes it a supplier question as much as a technology one.

Start with where the data goes. The model page lists regions in China, Singapore, Germany, the US, Japan and Hong Kong. The Australian Privacy Principles include APP 8, which covers the steps a business must take to protect personal information before it is disclosed overseas, and the OAIC says privacy obligations apply both to personal information put into an AI system and to AI output that contains it. Your privacy policy must also say whether you are likely to disclose personal information to overseas recipients.

Then test it properly. The OAIC says a business adopting a commercial AI product should check whether it has been tested for the intended use, how human oversight is built in, the privacy and security risks, and who can see the information going in and coming out. A side-by-side trial on your real tasks answers the first question; the others need answers from your supplier in writing. If a tool built on it helps sort job applications, check what candidates claim against verified records; a recruitment and candidate verification platform such as Talent Tabloid is built for that.

Agents need extra care whichever model they run on. After the Cohere finding, Cohere co-founder Ivan Zhang told The Register that the controls are largely the same whatever the cause: restrict unnecessary outbound access, monitor tool use, and do not treat a model's reasoning trace as a security boundary.

A short checklist:

  • List any tools or projects that use Alibaba's Qwen models, directly or through a supplier.
  • Ask which Alibaba Cloud region processes your requests, and whether personal information is involved.
  • Run a sample of real tasks on Qwen4-Max and on your current model, and compare the cost of each finished task, not the token price.
  • Give any agent only the access it needs, block outbound connections it does not need and log every action.
  • Update your privacy policy before customer information is sent overseas.

If you are weighing which AI model to build on, our IT consulting service can help you compare the options, and business systems and integrations covers connecting a model safely to the tools you already use. Agent controls and monitoring sit within our cyber security work. For a free first consultation and a written quote before any work starts, tell us about your project.

Frequently asked questions

What is Qwen4-Max?

Qwen4-Max is Alibaba's new flagship AI model, released on 16 November 2026. It reads text, images and video, has a one-million-token context window and is built for coding, office work and agent tasks.

How much does Qwen4-Max cost?

Alibaba Cloud's pricing page lists $1.5 per million input tokens and $4.5 per million output tokens in Singapore (International), in US dollars. Cached input costs $0.075 per million tokens, and a free quota of 1 million tokens for 90 days is offered in Singapore only.

Is Qwen4-Max really 60% cheaper?

Alibaba says the average cost per task fell by around 60% against Qwen3.8-Max on Work360, its own benchmark. Nobody has checked that figure independently yet, so measure the cost of your own tasks before relying on it.

Will there be open-weight Qwen4 models?

Alibaba says smaller, open-weight models from the Qwen4 series will soon be released to developers worldwide. Its release gives no date, no sizes and no licence terms for them.

Where is data sent to Qwen4-Max processed?

The model page lists regions in Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong. Ask your supplier which region handles your requests, and check your obligations under APP 8 before personal information goes overseas.

Key takeaways

  • Alibaba released Qwen4-Max on 16 November 2026, after a preview that opened on 11 November.
  • Its list price is $1.5 input and $4.5 output per million tokens, below Qwen3.8-Max's $2 and $6.
  • The around 60% lower cost per task is Alibaba's own benchmark result: test it on your own work.
  • Check which region processes your data, and your APP 8 obligations, before personal information goes in.
  • Agents on any model need tight limits on what they can reach and a log of what they do.

Need help with your business technology?

Tell us what you need. We aim to reply quickly.

Get a free quote