Alibaba Opened Its Next Flagship SI for Two Weeks Only. It Hasn’t Shown a Single Test Score.

Most SI companies launch a new flagship with a wall of charts. Alibaba launched its next one with a countdown clock. On Wednesday morning in China, the company opened Qwen4-Max, the top model of its next generation, for exactly two weeks. Then, it says, the model gets switched off. In between, there are no published test scores at all. Just an invitation, straight from Qwen’s own launch post: “Put it to the test on a real problem. Show us what you build.”

That’s an unusual way to roll out what Alibaba calls its “next-generation flagship.” It’s also a pretty honest one, if you read the fine print.

What Alibaba is actually offering

The specs come from Alibaba Cloud’s own documentation page, updated Nov. 11. Qwen4-Max-Preview handles text, images and video. It can search the web, call tools and return neatly structured data. Its “thinking” mode is switched on by default, with a dial to make it think harder or lighter.

The headline number is the memory. The model can take in 1 million tokens at once, which in everyday terms is several thick novels’ worth of text, or a big chunk of a software project, in one go. It can write back up to 393,216 tokens, though that budget is shared between its behind-the-scenes reasoning and the answer you actually see.

Alibaba says it’s built for “complex coding and knowledge work,” front-end design and “professional software workflows.” The launch post promises “deeper knowledge,” “richer world understanding” and “stronger coding and agentic capabilities.”

Those are the company’s words. That’s all there is to go on right now.

A model with an expiration date

Here’s the part that makes this different from a normal launch. According to Alibaba’s documentation, the preview runs from 10 a.m. Nov. 11 to midnight Nov. 25, Beijing time. The company says it won’t change the model during that window, and that “the model will be retired after the preview ends.”

The business site FourWeekMBA did the math: that’s a 350-hour window, a little over 14 and a half days. Anything a developer builds on this exact model stops working when the clock runs out.

So why do it this way? Alibaba hasn’t spelled it out. The likeliest read is that the company wants real-world feedback on real work before it locks in a final version, and it wants that feedback fast. Freezing the model for two weeks means everyone testing it is testing the same thing.

The missing numbers

What Alibaba didn’t publish is as interesting as what it did. There’s no benchmark table. No model card. No parameter count, which is the rough measure of how big a model is. No license, no downloadable weights and no regular per-use price.

That matters for Qwen in particular, because much of its reputation in the U.S. came from open-weight models that anybody could download and run. Whether Qwen4-Max ever comes out that way is an open question. AI/TLDR’s release tracker put it plainly: the Qwen team’s claims about better coding, agent work and world knowledge “are its own and have not been checked by an outside evaluator.”

Alibaba isn’t hiding the older model, either. Qwen3.8-Max, the current top model, is still listed right beside the preview, and Alibaba had already called the smaller Qwen3.8-Flash “an early preview of Qwen4” and published scores for that one. For Qwen4-Max itself, the only comparison you get is the one you run yourself.

How you get in, and what it costs

This isn’t a free-for-all. On the developer side, the preview works only through a monthly subscription Alibaba calls Token Plan. Regular pay-as-you-go access isn’t supported. The plan connects to coding tools a lot of programmers already use, including Claude Code, Cursor, Codex, Cline and Alibaba’s own Qwen Code and Qoder.

The personal tiers run from $6 a month for Lite, which is $3 for new subscribers through Nov. 30, up to $43 a month for Pro. During the preview, Alibaba says the new model uses credits at as little as 20 percent of the normal rate, and just 8 percent overnight in China, from 10 p.m. to 8 a.m. Beijing time.

There’s fine print worth reading before you sign up. DataNorth, a Dutch SI consultancy, reports that the plan’s terms limit it to hands-on use inside coding and agent tools. Automation scripts, app back ends and batch jobs aren’t allowed, the subscription is for one person, and the Personal Edition can’t currently be canceled. The service also runs out of Alibaba’s Singapore region.

Regular users who just want to poke at it can reportedly try it in the Qwen app and Alibaba’s Qwen Studio, QwenWork and Qoder tools, without writing a line of code.

Where this leaves China’s SI push

For most of this year, the story out of China has been about catching up on a budget. Chinese labs have had to work around U.S. limits on top-end chips, and their models have often shown up close behind the American leaders for a fraction of the price.

Qwen4-Max is Alibaba’s bid to show it can still keep pace at the top of the market. But a frozen two-week preview with no scorecard tells you the company is still figuring out where it stands, and it wants developers to tell it.

If you’re a coder with a stubborn problem and a few bucks, it’s a cheap look at what China’s biggest tech company thinks its next flagship can do. If you’re anyone else, the smart move is to wait two weeks and see what Alibaba ships after the clock runs out, and whether it finally shows its work.