Alibaba has opened a preview of Qwen4-Max, the first model of its new Qwen4 generation, giving developers their first chance to try the company's next flagship artificial intelligence model. The preview went live on 11 November on Alibaba's Token Plan subscription, but it comes with a hard expiry date: the model will be retired on 25 November, just 15 days after it arrived.
The announcement came from the official Qwen account on X, which posted simply that "Qwen4 is coming", and from Alibaba Cloud's Model Studio documentation, which lists the model under the identifier qwen4-max-preview. It is available through the Token Plan, a credits-based monthly subscription for coding tools, and in early access in Alibaba's own apps, including QwenWork, Qoder, the Qwen app and QwenStudio.
What Alibaba has published about the model is narrow but concrete. It has a context window of 1,000,000 tokens, and a maximum output of 393,216 tokens, a ceiling that covers the model's reasoning and its final answer together rather than the answer alone. It accepts text, image and video input, and its thinking mode is on by default. The Token Plan page lists three capability tags for the model: text generation, reasoning model and visual understanding.
Access is deliberately limited. The model is available through the Token Plan only, with no pay-as-you-go option, and an ordinary Model Studio key does not open it. For new users, the Token Plan's Lite plan starts at 3 dollars a month as a promotion, and the model is flagged with a limited-time 80 percent discount on credit usage. The preview runs from 10:00 on 11 November to 24:00 on 25 November, Beijing time, and Alibaba says no changes will be made during that period.
What Alibaba has not published is just as striking. There is no parameter count, no model card, no published benchmark score, no per-token price, no open weights and no licence. The company has not said when the full Qwen4 release will arrive, or when the other tiers of the family, including Plus, Flash and the 27-billion-parameter model, will ship. As of the preview's launch, Alibaba Cloud's model marketplace had no Qwen4-Max listing for pay-per-token use.
The preview is the first Qwen4 product anyone can actually call. Alibaba previewed the Qwen4 family at its Apsara conference in Hangzhou in September, where Qwen project lead Liu Dayiheng said the family was in training on a new-generation architecture. The only other part of the family anyone can download is Qwen3.8-Flash-Next, released in late August with open weights.
The launch lands in a crowded week for AI. On the same day, Google DeepMind introduced AlphaGenome 2, a model for predicting how DNA changes act in specific cell types, and Runway released the open weights of its Praxis-1 robot model. Anthropic unveiled its Claude Fable 5.2 flagship days earlier, and OpenAI reported that its quarterly revenue had topped 10 billion dollars for the first time. Alibaba's move is a reminder that the race for the next generation of models is being fought on several fronts at once, and that China intends to stay in it.
For developers, the practical question is what the preview actually does. Early testers have reported mixed results, with one account of a 300-token reply that contained nothing at all, and no independent benchmark of the preview existed as of 11 November. Without published scores, the model's capabilities can only be judged by trying it, which is precisely what the 15-day window invites developers to do.
The short window is itself a statement. By putting the flagship behind a subscription, with a fixed retirement date and no price list, Alibaba is treating the preview as a controlled test rather than a launch. Developers who want to use it will have to move quickly, and whatever they build on it will need to be rebuilt when the preview model disappears on 25 November.
Whether Qwen4-Max lives up to its billing as a flagship will only become clear when the full family ships, with benchmarks and a price. For now, the preview tells the industry two things: that Alibaba's next generation is real enough to touch, and that the company is in no hurry to give it away.
