Alibaba released Qwen4-Max on November 16, moving its top model into regular hosted access after a limited preview began on November 11. The launch gives software teams a new option for work that combines long instructions, documents and visual material in one request.
Alibaba Cloud said the model is available via the QwenCloud API and Token Plan and through Alibaba products including QwenWork and QwenStudio. Alibaba Cloud Model Studio lists the production model as qwen4-max. The company’s Token Plan product updates give the earlier preview a separate identifier and a November 25 end date, a distinction for teams that configured an integration during the preview.
What the model can take in
Model Studio lists text, image and video as supported inputs, but the output is text. That means a developer can ask for a description of a video clip or structured information from an image, but should not treat Qwen4-Max as a video generator. The API also supports structured responses and tool calls, which let an application request output in a predictable format or give the model access to functions the developer provides.
Alibaba puts the context window at one million tokens. Tokens are the units a model uses to process input; the published window describes capacity, not a promise that the model will catch every detail. Model Studio lists a lower maximum input length when thinking mode is on, and a separate combined limit for generated reasoning and the final answer. A team assessing long files should check how many requests fit comfortably under those limits, rather than simply comparing the headline window size.
Prices differ by deployment
Model Studio lists Singapore’s international deployment at $1.50 per million input tokens and $4.50 per million output tokens. Its global deployments in Frankfurt, Virginia, Tokyo and Hong Kong list $1.413 for input and $4.24 for output at the same unit. The launch post also advertises a lower price for an input cache hit, which applies when eligible material can be reused. That is not the price for a new request’s full input.
Alibaba says the model has 2.4 trillion parameters in total and activates roughly 71 billion for each token. This is a mixture-of-experts design: the system routes work to only part of the model at a time. Alibaba reported lower token use and cost per task than Qwen3.8-Max on its own Work360 evaluation. Those figures reflect Alibaba’s task mix and settings, not an independent measurement of savings for customers.
What to watch next
Alibaba says it plans to release smaller Qwen4 models with open weights, the numerical settings needed to run a model outside a vendor’s hosted service. The launch announcement describes hosted access for the flagship and gives no release dates for those smaller models. For teams deciding whether to switch now, the useful comparison is the cost and accuracy of completing their own tasks on the paid endpoint, particularly when outputs need correction or review.



