Alibaba has opened Qwen4-Max-Preview on 11 November 2026, with a preview available until 25 November. For us, this window justifies a targeted trial on the development of websites and business tools, but its announced withdrawal rules out making it a lasting dependency for your company.
Qwen4-Max-Preview: what changes and until when
Alibaba Cloud’s documentation presents a model able to process text, images and video, intended in particular for development and agent tasks. It supports tool calls, which make it possible to invoke external functions, and structured outputs, which software can use in an imposed format.
Another announced feature: a context window of one million tokens, the units of text processed by the model. For a development trial, we would look at how it uses a set of files and instructions. The criterion would be its ability to find the right information and respect the constraints of the project.
Access in Model Studio is exclusively through the Token Plan, with no pay-as-you-go option. Alibaba states that the preview will remain unchanged during the trial period, then will be withdrawn after 25 November. This deadline should set the scope of the test from the start.
What this means for an SME in French-speaking Switzerland

If you run an SME, the practical interest lies in the tasks to entrust to the model. We would pick, for example, analysing a screenshot to prepare an interface fix, proposing a patch on a copy of a site or preparing tests for a form. Each exercise should lead to a result that we can check.
For your company, the useful question would be: how much work remains after the answer? A code proposal deserves to be assessed together with its review, its corrections and its validation. We would also count the errors introduced and the instructions forgotten. A quick answer that requires lengthy rework would add little to the project.
Access deserves separate attention. Alibaba presents the Token Plan as a subscription for interactive use in compatible AI tools. Our starting point would therefore be a development assistant used by one person, within a defined scope. Before considering any other use, we would check the conditions that apply to the account and the tool chosen.
Our view: testing with an end date
At Smart Impact, our position is clear: a temporary preview can be used to learn, without becoming a foundation for a customer service. The test must produce a useful decision even when access to the model stops. We want to come away with well-described tasks, quality criteria and comparable results.
This changes the way the trial is prepared. We would favour demonstration files, an environment separate from the public site and human validation before any change is kept. For tool calls, we would limit the available actions to the strict minimum: reading test files, proposing a change, running a supervised check.
The large context interests us above all as an object of evaluation. We would ask the model to justify its proposals from the material provided, then check those references. Piling documents into a request would be of little value if the answer overlooks a business rule that matters for your activity.
This is the approach we advocate for AI and automation: start from a defined task, measure the work actually accomplished and keep control of what follows. Here, the withdrawal date makes this last point particularly concrete.
A useful trial in three points
-
Choose a task and its expected result. We would prepare a representative exercise, such as fixing a display defect on a copy of a page. The instructions would specify what must work at the end and what must be preserved. The same exercise would be used to compare the answers.
-
Check access and supervise actions. Before the test, we would check the uses permitted by the Token Plan, the compatible tools and the data required. Permissions would remain limited to the scope of the exercise, with a review of the proposals before they are applied.
-
Finish before the withdrawal. The review would bring together the rework time, the errors and the quality of the validated result. We would keep the instructions, the tests and the conclusions for a future evaluation. No day-to-day operation of the site or of a business tool should depend on this temporary access.




