Alibaba has put a preview of its next flagship model, Qwen4-Max-Preview, in front of developers, but only for two weeks. It runs from 11 November to 25 November 2026, can be reached only through Alibaba Cloud’s Token Plan subscription, and will be retired when the window closes. Pay-as-you-go access is not offered.
The model reads text, images and video and handles a context of 1 million tokens, with up to 393,216 tokens of output. That output limit is shared between the model’s reasoning and its final answer, so a long chain of thought eats into the space left for the reply. Alibaba pitches it at programmers, office work, web design and multi-step agents.
That sharing is the main thing existing users need to know. With this model, the max_tokens setting in the OpenAI-style chat API counts reasoning and answer together, where older Qwen models keep their previous behaviour. A limit that used to size only the answer can now cut the reply short, so apps moved across need more headroom.
Reasoning is on by default, at the highest of three levels: low, medium or xhigh. The same call should not set both a reasoning effort and a thinking budget. In multi-turn chats the earlier reasoning must be sent back in a separate field and not folded into the message text. The model supports function calling, web search, built-in tools and JSON output, and works with the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats.
To try it, a developer picks Token Plan Personal or Token Plan Team, creates the plan’s own API key and points the client at the dedicated endpoint in Alibaba’s model documentation, with the model ID qwen4-max-preview. Alibaba says nothing will change during the preview.
Qwen is already a name many readers have met: the family also powers part of Apple Intelligence in China. The preview is Alibaba’s early look at the family’s next top-end model.
Image and video input suits jobs such as reading a screenshot, working through a document or analysing a clip. Structured output works with both JSON Object and JSON Schema formats. Setup instructions, the list of compatible tools and the credits bundled with each plan sit in separate Token Plan guides.
The page gives no benchmark scores, parameter count or downloadable weights, and no per-token price, so there is nothing yet to compare against rival models. Developers who want to test it have until 24:00 on 25 November, Beijing time (UTC+8).
