Alibaba is using its first Qwen4 release to sell Token Plan subscriptions, and its developer feedback window closes before the full Qwen4 family's undated launch.
Alibaba opened a preview of Qwen4-Max on November 11, the first model of the next-generation Qwen4 family, and limited it to subscribers of its Token Plan for 15 days. The model accepts text, image and video input and handles up to one million tokens of context, TechNode reported. Alibaba’s QwenCloud documentation says the preview runs from 10:00 on November 11 to 24:00 on November 25, Singapore and Hong Kong time, after which the model will be retired, and that pay-as-you-go access is not supported.
Alibaba also said the full Qwen4 generation is "coming soon", promising deeper knowledge, richer perception of the world and further gains in coding and agent capabilities, IT之家 reported. The new model is already built into Alibaba’s own products: QwenWork, the Qoder coding tool, the Qwen app and QwenStudio.
Alibaba has published no benchmark results or comparisons with other models: none appear in the coverage, including that of Phoenix News, which attributed its details only to Alibaba, and the preview documentation gives no parameter count, per-token price or download of model weights. For now, Qwen4-Max is a product preview rather than a fully documented model release.
The preview is built for long coding and agent work
The documentation describes Qwen4-Max-Preview as the Qwen family’s next-generation flagship multimodal model, combining text generation, image and video understanding and deep reasoning. It is aimed at coding, office productivity, front-end design and complex agentic tasks, and it supports function calling, web search, built-in tools and structured output. Developers can reach it through interfaces compatible with OpenAI’s Chat Completions and Responses formats and with Anthropic’s Messages format. Phoenix News added 3D modeling and the operation of professional software to the list of target uses.
Reasoning is switched on by default. Users can set the reasoning effort to low, medium or xhigh, and xhigh is the default. Qwen4-Max-Preview can return up to 393,216 tokens in a single response, a budget shared between its reasoning and its final answer. The model ID is qwen4-max-preview, and bills show it as qwen4.0-max-preview, according to the Token Plan subscription page.
Alibaba’s line-up already offered a window of that size. Its model changelog shows that Qwen3.8-Max, listed on August 3 with 2.4 trillion parameters in a mixture-of-experts design, already offered a one-million-token context. The step up in Qwen4-Max is in video input, agent work and reasoning depth, none of which Alibaba has yet quantified.
Token Plan pricing turns the launch into a subscription drive
Routing the preview only through Token Plan, Alibaba’s monthly subscription for model usage, ties the new model to recurring revenue. On the international QwenCloud site, individual plans cost US$6 a month for Lite, US$10 for Essential, US$18 for Standard and US$43 for Pro, with new users paying US$3 and US$5 for the two cheapest tiers until November 30. During the preview, calls to Qwen4-Max consume as little as 20% of the standard credit rate, falling to 8% between 22:00 and 08:00.
In China, the Lite and Essential personal plans start at 39 yuan a month at the current limited-time price, and new customers get a further 50% off, bringing the entry price to 19.5 yuan, IT之家 reported. Phoenix News described the stacked offer as a 20% rate for Token Plan users with an extra half off for newcomers, or as little as 10% of the original price.
The preview window falls in the same weeks as the new-user discounts, which end on November 30. A model that expires after two weeks gives developers a reason to subscribe now, and gives Alibaba usage from paying customers before the general release.
Qwen4’s full release is the next marker
Alibaba set out a broader Qwen roadmap at its Apsara conference in September, saying Qwen4 was in training and that later Qwen 4.5 and Qwen 5 models were projected at 5 trillion to 10 trillion parameters, as EastFrontier reported at the time. In August it had also said it would open-source Qwen3.8-Flash-Next as an early look at the Qwen4 architecture.
The November 25 cutoff is the date to watch. Alibaba will have to decide whether a stable Qwen4-Max replaces the preview, whether it comes with a benchmark table and per-token API pricing, and whether smaller Qwen4 models follow with open weights. Independent test results on coding and agent tasks would show whether the gains Alibaba describes hold up outside its own products.