The Clock Is Already Ticking on Qwen4-Max
Picture this: you finally get your hands on a brand-new flagship AI model, you start testing it, and then you notice the fine print. It disappears in fourteen days.
That is what Alibaba has done with Qwen4-Max. The Chinese tech giant unveiled its next-generation AI model family on Tuesday, November 11, and opened a preview of its flagship model that morning. But this is no ordinary launch. The model comes with a hard expiry date, and the countdown started the moment it went live.
Alibaba announced the Qwen4 family through its official Qwen account on X, and the flagship Qwen4-Max preview appeared on Alibaba Cloud’s Token Plan at 10:00 a.m. Beijing time, according to reports. Within hours it was also reachable inside Qwen Chat, the company’s consumer app.
For a model family that was only teased five days earlier, the move is bold. For the developers and AI watchers trying to judge it, the move is also deeply frustrating. Alibaba has shared almost nothing about what is under the hood.
What Alibaba Actually Announced With Qwen4-Max
The announcement itself was short on detail and long on promise. Alibaba said Qwen4 is its next-generation model family and that Qwen4-Max is the flagship tier, built for stronger coding and more capable AI agents.
The preview is available in five places: the Token Plan subscription on Alibaba Cloud’s Model Studio, QwenWork, the coding tool Qoder, the Qwen app, and QwenStudio. The model identifier is qwen4-max-preview, and it accepts text, images and video as input.
What Alibaba did not publish is just as telling. There is no model card, no parameter count, no benchmark scores and no per-token price. There is also no word on open weights, the feature that made earlier Qwen releases so popular with developers around the world.
Alibaba had named four Qwen4 tiers back in September at its Apsara conference: Max, Flash, Plus and 27B. Only Max has appeared so far, and only as a preview.
Qwen4-Max Specs: A Million Tokens of Context
The one hard number Alibaba did share is impressive. Qwen4-Max supports a context window of one million tokens, with a maximum output of 393,216 tokens that the model’s thinking and its final answer have to share.
In plain terms, that means the model can hold an enormous amount of material in its working memory at once: whole codebases, long documents, hours of meeting transcripts. It also means it can produce very long outputs, from detailed reports to sprawling pieces of code.
The model also supports reasoning, tool calls, structured outputs and built-in tools such as web search. Alibaba’s documentation lists screenshot analysis, document understanding and video analysis among its intended uses.
For comparison, the model it replaces, Qwen3.8-Max, has been generally available since August 3 with the same one-million-token context, priced at $2 per million input tokens and $6 per million output tokens. Qwen4-Max has no published per-token price at all.
Five Ways to Try Qwen4-Max — and the Catch
Getting access is easy, but only on Alibaba’s terms. The Token Plan is a monthly subscription based on credits, and new users can start from $3 a month on the Lite plan as a promotion. The model is flagged with a time-limited 80% discount on credits.
Qwen Chat offers a second door. The preview turned up in the chat product on the same day, so anyone with an account can open a browser tab and start typing. No API key, no subscription, no per-token meter.
But the access comes with walls. The Token Plan route uses a dedicated base URL, so a Model Studio key does not open it. Pay-as-you-go billing is explicitly not supported. And the plan’s terms restrict the model to interactive use inside coding and agent tools, so you cannot build it into an application backend.
That is a deliberate choice. Alibaba is offering a test drive, not a product you can build a business on. At least, not yet.
The 14-Day Expiry Date on Qwen4-Max
Here is the strangest part of the whole launch. The Qwen4-Max preview runs from 10:00 a.m. on November 11 to midnight on November 25, Beijing time. After that, Alibaba says the model will be retired.
Fourteen days. Then it is gone. Alibaba also says the model will remain unchanged during that window, which turns the preview into a fixed evaluation target. Any result you gather describes this exact version, and only this version.
It is a sharp contrast with how rival labs launch flagship models. When Google released Gemini 4 Carbon this week, it arrived with benchmark scores, pricing and a general-availability date. When Anthropic launched Claude Fable 5.2, it shipped to paying customers the same day.
Alibaba, by comparison, has given the world two weeks to look, and no promise about what comes after. The full Qwen4 series has not been officially released. The preview may be a teaser for a bigger launch, or a quiet test of demand before the real thing.
What Is Missing: Benchmarks, Weights and a Price
The gaps in the announcement matter, because they decide whether Qwen4-Max is a genuine contender or just a headline. As of November 11, there are no benchmark scores from Alibaba and none from independent testers.
Artificial Analysis, the third-party leaderboard many developers treat as a neutral reference, carries no Qwen4 entry at all. That means there is no intelligence index score, no cost-per-task figure and no speed measurement to compare against rivals like GPT-6.1 Sol or Gemini 4 Carbon.
There is also no word on downloadable weights or a licence. Earlier Qwen models won fans precisely because developers could download and run them. A closed, hosted preview that vanishes in two weeks is the opposite of that tradition.
And then there is the price. The Token Plan meters the model in credits, and Alibaba does not publish a credits-per-token rate. Without that, nobody can say whether Qwen4-Max is cheap or expensive for real work.
Why the Qwen4-Max Launch Matters Anyway
Even with all those caveats, this launch is worth watching. Qwen is one of the most widely used open model families in the world, and its flagship sets the pace for a whole ecosystem of Chinese AI products, from coding tools to agent platforms.
The timing is also significant. November has become a crowded month for AI launches. Google has shipped Gemini 4 Carbon and Gemini Robotics 2.5, Anthropic has launched Claude Fable 5.2 and a full office suite, and OpenAI has rolled out its GPT-6 family. Alibaba clearly did not want to sit this month out.
The preview format may also be a strategy. By putting the model in front of users for a short, fixed window, Alibaba gets real-world feedback and usage data without committing to a price, a service level or a permanent product. If the feedback is strong, the full launch follows.
There is a risk in that approach, though. Developers who invest time testing a model that disappears in two weeks may not come back for the next one. Trust, in the model business, is built on stability.
What Happens Next for Qwen4-Max
The most likely next step is a full Qwen4 launch once the preview window closes on November 25. Alibaba has already named the other tiers, Max, Flash, Plus and 27B, and the preview is clearly a stepping stone rather than the destination.
Watch for three things in the coming days. First, independent benchmark results, as testers push the model through coding and reasoning tasks. Second, any statement from Alibaba about open weights, which would decide whether developers embrace Qwen4 the way they embraced Qwen3.
Until then, the clock is ticking. If you want to try Qwen4-Max, you have until November 25. After that, the model retires, and the only thing left will be the questions it raised.
For now, Alibaba has given the AI world a tantalising glimpse of its next flagship, and a deadline to judge it by. Fourteen days is not long, but in a month this crowded, it may be all the time Qwen4-Max needs to make its case, or to be forgotten.
