Good evening, everyone.
The model everyone has been waiting for finally showed up on Tuesday, and Alibaba left out the numbers.
Qwen 4 is here. Or half of it is. Qwen4-Max Preview opened on November 11, three days after the window Alibaba's own chief technology officer had pointed at when he said "next week" on October 29. It has a million-token memory, it takes text, images and video, and it will be switched off on November 25. That is the whole story in one sentence, and it tells you almost everything about how this launch works.
Here is my position, up front. Call it a two-week try-out with a subscription form attached, because that is what it is. A flagship model with no benchmarks, no price and no weights is not a product you can judge. It is a product you can only pay to peek at.
What you actually get
The spec sheet is short, and the shortness is the point. Qwen4-Max Preview reads up to a million tokens at once, which is a lot of room. A whole codebase, a stack of papers, a long video transcript, all in one request. It can write back up to 393,216 tokens in one go, though that number covers its own thinking as well as its final answer, so the answer you see is shorter than the number sounds.
It is a multimodal reasoning model, built for hard coding, knowledge work and front-end builds, according to Alibaba. You can try it in four places: QwenWork, Qoder, the Qwen App and QwenStudio. Developers can call it through an API, with the model ID qwen4-max-preview.
That is the full list. There is nothing else on the sheet.
What you do not get
- Benchmark scores: none, from Alibaba or anyone else. It is not on the Artificial Analysis leaderboard.
- Parameter count: not published, and no model card saying what it can and cannot do.
- Per-token price: none. There is no pay-as-you-go option at all.
- Weights: none to download, and no licence, so nothing to run yourself or fine-tune.
- Staying power: none promised. The preview runs from November 11 to November 25, and then Alibaba retires it.
Five missing pieces, and each one is the kind of thing a normal flagship launch puts on the first page.
A preview is not a launch
The only way in is a paid Token Plan subscription. You sign up, you buy credits from a monthly quota, and you get two weeks with a model that has no public scores. That is a funnel. A release says here is what this is, here is what it costs, go build on it. A funnel says pay first, find out later, before it disappears in fourteen days.
And you cannot build on it, even if you wanted to. A model with an expiry date is useless for real work. Anything you wire into your app or your scripts dies on November 25. The other three tiers, Plus, Flash and a 27-billion-parameter model, have no date at all. So the family is still a promise, and the one piece you can touch is a rental.
Alibaba first showed the Qwen 4 family at its Apsara conference in September, then said nothing until Tuesday. The contrast with this week is what makes it sting. Anthropic shipped Fable 5.2 on Monday with real benchmark scores and a price cut. Google put Gemini 4 Carbon on general availability with a price on the page. NVIDIA shipped its robot brain with numbers, even if the numbers were picked. Qwen shipped a subscription form.
The only public signal so far comes from outside Alibaba. For four days before the announcement, Qoder users were quietly testing an anonymous model called QoDuck, which turned out to be this preview. One community leaderboard for game-feel tests put it at 10.3, below Qwen3.8 Flash and Max. That is one small test family, from one community board, on a build that changed daily. It proves nothing, and it is still the most anyone outside Alibaba can point to.
That is the honest state of it. The model might be great. The model might be ordinary. Nobody outside Alibaba can say, because Alibaba published nothing you can measure.
So who should care? If you are curious, try it in QwenStudio and see how it feels on your own work. That is a fair use of a preview. If you build products, wait. Wait for the real release, with a model card, a price and a date that does not come with an expiry.
I wrote on Monday that the whole internet was waiting for Qwen 4. That wait is over now. It arrived as a try-out, and the difference matters more than the arrival.
Would you pay for a subscription to try a model that disappears in two weeks, or would you wait for the real thing? I am curious where people land on this one.