Alibaba opens a Qwen4-Max preview with a 1 million-token window, and the English page retires it on 25 November
The Model Studio page, updated Wednesday, limits the model to Token Plan, sets thinking to xhigh by default, and says max_tokens now counts the reasoning. A verified Qwen post the same morning names Qoder and the Qwen app, and a discount the docs do not print.
Feature2 minIn-depth
James Okonkwo2 min read
- Qwen
- Alibaba
- Open weights
- Agents

Photo: Qwen, via SinaSource page
Alibaba on Wednesday opened qwen4-max-preview, which its Model Studio page calls the next-generation flagship multimodal model in the Qwen family. The page was last updated on 11 November. The preview runs from 10:00 on 11 November to 24:00 on 25 November, China time. The page says no changes will be made during that window, and that the model will be retired after it. Pay-as-you-go is not supported. Access is through Token Plan only, on a base URL in ap-southeast-1, with the model id qwen4-max-preview. TechNode’s note the same day is shorter: a preview of Qwen4-Max, image and video inputs, up to a million tokens of context, on Token Plan.
The context length on the English page is 1 million tokens. The maximum output is 393,216 tokens, shared by the thinking trace and the final answer. Thinking is on by default. The effort settings are low, medium, and xhigh, and xhigh is the default. The page says not to set reasoning_effort and thinking_budget together. From this model on, max_tokens is equivalent to max_completion_tokens, and both count thinking plus the answer. A limit sized for the answer alone can cut the reply off. Earlier Qwen models, the page says, keep their previous behavior. Chat Completions also turns preserve_thinking on by default, and the instructions say to return historical thinking in reasoning_content rather than fold it into content.
What the call is allowed to do, and where else the name appears
The capability table is text and reasoning, image and video understanding, tool calling, and structured output. Tool calling covers function calling, web search, and built-in tools. Structured output covers a JSON object and a JSON schema. The page says the model is compatible with OpenAI’s Chat Completions and Responses APIs and with Anthropic’s Messages API. The uses it names are coding, office work, frontend design, and agent tasks. It does not print a price per token, a parameter count, or a benchmark.
A post from the verified Qwen account, carried on Sina and timestamped 10:09 on 11 November, is the place those missing commercial lines sit. It says a new Qwen4 generation is about to be released, and that a Qwen4-Max preview is already on Alibaba’s Token Plan, at a limited-time rate as low as one-tenth of the usual price. It also says the model is in 千问办公, Qoder, the Qwen app, and QwenStudio. TechNode’s English list for the same early access is the Qwen app, Qoder, QwenWork, and QwenStudio. The Model Studio page names none of those products, and it does not mention a tenth-price rate. The account’s post describes results on coding, office work, frontend design, and software operation as striking. It does not attach a score.
The English page’s last operational fact is the end of the window. After 25 November at midnight, China time, the preview id is scheduled to come off. Until then the page says the model will not change. A developer who calls it is calling a checkpoint with a removal date, a shared budget for thinking and answer, and no public price list outside the Token Plan the account says is discounted.


