Alibaba’s Qwen4-Max Preview Has a November 25 Deadline | Zeniteq
Loading
Loading article
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
Alibaba’s Qwen4-Max Preview Has a November 25 Deadline
Qwen4-Max-Preview brings a million-token context window, but developers need a Singapore-region Token Plan, separate credentials, and an exit plan before retirement.
Alibaba’s Qwen4-Max-Preview is a time-limited multimodal model available until November 25, 2026. Developers must secure a Singapore-region Token Plan subscription and use specific sk-sp- credentials with dedicated non-regional endpoints to access its million-token context window. Because reasoning and final answers share a single token budget, developers should adjust output limits to prevent truncation.
This preview requires careful integration, as it mandates specific handling of reasoning content in multi-turn conversations. Since the model will be retired immediately after the deadline, teams must treat this as a strictly time-boxed evaluation rather than a production-ready deployment.
Alibaba’s QwenCloud released Qwen4-Max-Preview on November 11, 2026, introducing what it calls the next-generation flagship multimodal Qwen model. The release supports text generation, image and video understanding, reasoning, tool calling, and structured outputs, with OpenAI- and Anthropic-compatible APIs. Access is limited to QwenCloud’s subscription-based Token Plan; pay-as-you-go accounts cannot call this preview. (Release changelog, model documentation)
The deadline matters more than the model name. QwenCloud says the preview will be retired after 24:00 on November 25, 2026, UTC+8. Its documentation also changes how output limits work: thinking and the final answer now consume the same token budget.
My view is that developers should treat this as a time-boxed compatibility and capability evaluation. The documented retirement gives any application built around this model a known expiration date, regardless of how well it performs.
The Exact Access Window
QwenCloud specifies the following schedule:
Opening: November 11, 2026, at 10:00 UTC+8, equivalent to 02:00 UTC.
Closing: November 25, 2026, at 24:00 UTC+8, equivalent to 16:00 UTC that day.
After closing: The preview model will be retired.
The closing time means the end of November 25 in UTC+8, or 00:00 on November 26. The full window is therefore 14 days and 14 hours, not exactly 14 days. QwenCloud says no changes will be made during the preview period. (Preview schedule)
That fixed window provides a useful evaluation boundary: developers can assess one documented preview without a planned mid-window model change. It also means a late start leaves less time to fix integration problems, compare reasoning settings, and save results before access ends.
The temporary element is Qwen4-Max-Preview’s availability. Token Plan itself is an ongoing subscription service covering multiple models and tools; the documentation does not describe the subscription as ending with this preview. (Token Plan overview)
QwenCloud positions Qwen4-Max-Preview for coding, office productivity, frontend design, and complex agent tasks. Its launch documentation lists these capabilities and limits:
Feature
Documented support
Model ID
qwen4-max-preview
Context window
1 million tokens
Maximum output
393,216 tokens, shared between thinking and the final answer
Inputs
Text, images, and video
Reasoning
Enabled by default
Tools
Function calling, web search, and built-in tools
Structured outputs
JSON Object and JSON Schema
Compatible APIs
OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages
A million-token context window can accommodate substantial source material, conversation history, and tool results. The practical question is whether the model reliably finds and uses the relevant information inside that material. Capacity alone cannot answer that.
The output ceiling needs similar care. 393,216 tokens is a combined reasoning-and-answer allowance, so it is not a promise of a final answer that long. Tokens are units of model processing, not a fixed number of words or pages.
The release documentation also describes image and video understanding. Developers can evaluate tasks such as screenshot interpretation, document analysis, or questions about video input. They should not interpret that capability list as an announcement of image or video generation.
Who Can Access the Preview
Developers need Token Plan Individual or Token Plan Team and the corresponding dedicated API credentials. Token Plan can be purchased and used only in QwenCloud’s Singapore region. When another region is selected in the console, the purchase entry is hidden. (Token Plan overview)
Singapore here refers to the service region selected in QwenCloud. The cited documentation does not establish a Singapore-residency requirement.
Token Plan is a subscription billed in Credits, intended for interactive use in compatible AI tools. A tool must support the relevant OpenAI or Anthropic protocol and allow a custom API key and Base URL. An interface that resembles a supported coding assistant is insufficient if its underlying requests use a different protocol.
The setup sequence is straightforward:
Switch the QwenCloud console to Singapore.
Obtain a Token Plan Individual or Team subscription.
Retrieve the dedicated Token Plan key from the appropriate tab on the API Keys page.
Token Plan keys start with sk-sp-. Regular pay-as-you-go keys typically start with sk-ws-, with older keys using sk-. The two credential types are separate and cannot be used interchangeably.
The Individual quickstart publishes these protocol-specific Base URLs:
The dedicated hostname contains no region ID. Do not substitute a regional pay-as-you-go hostname such as ap-southeast-1.maas.qwencloudapi.com. (Endpoint and key setup)
Changing only the model ID is therefore insufficient. A client configured for an existing Qwen pay-as-you-go service needs a different key and Base URL too.
There are two distinct failure risks. Mixing key and endpoint types is unsupported. Separately, valid calls to other models through pay-as-you-go credentials remain billed independently, even when the account has a Token Plan subscription. QwenCloud explicitly says that owning a subscription does not automatically cover those calls. (Authentication and billing rules)
I’d keep the preview configuration separate from the normal application configuration. That makes accidental credential mixing easier to spot and makes retirement cleanup a configuration change instead of a search through source code.
The Output Budget Change Can Truncate Answers
For OpenAI Chat Completions, QwenCloud introduces a consequential change with qwen4-max-preview: max_tokens is equivalent to max_completion_tokens. Both limit the combined tokens used by thinking and the final answer. Existing Qwen models retain their previous behavior.
The documentation warns:
“Thinking consumes part of the same budget, so a limit previously sized for the answer alone may truncate output.”
An application that previously allocated 8,192 tokens for its visible answer should not assume the same setting preserves that answer allowance. As a hypothetical example, if thinking consumes 6,000 tokens from an 8,192-token combined budget, at most 2,192 tokens remain for the final answer. That arithmetic illustrates the documented mechanism; it is not a measured preview result.
An illustrative Chat Completions request body could look like this:
{
"model": "qwen4-max-preview",
"messages": [
{
"role": "user",
"content": "Explain the main risks in this proposed API migration."
}
],
"reasoning_effort": "low",
"max_completion_tokens": 8192
}
This example selects the preview, requests low reasoning effort, and caps the combined output at 8,192 tokens. It has not been executed, and that limit is an example, not a recommended value for every workload.
QwenCloud lists low, medium, and xhigh reasoning effort, with xhigh as the default. Thinking is enabled by default, and the documentation says not to set reasoning_effort and thinking_budget together. (Reasoning settings)
A useful evaluation should vary reasoning effort and output budget separately. Otherwise, an incomplete answer could be mistaken for poor task performance when the request simply ran out of its shared allowance.
Multi-Turn History and Built-In Tools Need Attention
Chat Completions enables preserve_thinking by default for this preview. QwenCloud instructs developers to return all historical thinking in assistant messages’ reasoning_content field and keep it separate from content. (Conversation handling)
That requirement deserves a client-level check. A conversation store that saves only visible assistant text may omit information the documented continuation workflow expects. Inspect the messages your application actually sends on later turns, including after tool calls.
Built-in tool access introduces another protocol distinction. The Token Plan quickstart says Harness tools, such as web search and code interpreter, are automatically invoked through the Responses API. A Chat Completions-only tool can produce ordinary responses but will not automatically invoke those Harness tools. (Harness integration notes)
Function calling and Responses-based Harness tools should therefore be separate evaluation items. Protocol compatibility does not mean every client exposes every capability.
A Migration Checklist With an Exit Plan
The most useful preview outcome is a reproducible record of behavior, configuration, and failure cases. Before committing evaluation time, work through these checks:
Confirm subscription access. Use the Singapore region and verify the exact qwen4-max-preview model ID.
Pair credentials correctly. Use an sk-sp- key with the matching Token Plan Base URL. Store the key server-side in an environment variable or secrets manager.
Check the client protocol. Confirm whether it uses Chat Completions, Responses, or Anthropic Messages, especially if built-in tools matter.
Revisit output limits. Allow for both reasoning and the final answer; log incomplete outputs while comparing reasoning settings.
Inspect conversation history. Verify that multi-turn Chat Completions requests retain reasoning_content separately from visible content.
Exercise real task constraints. Check schema validity, tool arguments, screenshot interpretation, and retrieval of relevant facts from long inputs.
Schedule retirement cleanup. Save evaluation results, remove the preview dependency, and restore an available model before the cutoff.
The credential, endpoint, budget, and history requirements come from QwenCloud’s documentation. The task checks and exit plan are practical recommendations derived from those requirements and the fixed retirement date. (Model usage notes, API key security guidance)
The launch pages cited here provide no full benchmark results, safety evaluation, architecture details, parameter count, or pay-as-you-go pricing for this preview. They also name no successor or general-availability date.
That leaves a clear boundary for adoption decisions. Developers can use the window to assess their own workloads, but the documented evidence is insufficient to rank Qwen4-Max-Preview against competing AI models or plan a lasting deployment around it. Retirement is certain in the published schedule; replacement availability is unspecified.
Frequently Asked Questions
3 questions
1
Can I Use My Existing Pay-As-You-Go Key?
No. Qwen4-Max-Preview requires Token Plan access, an sk-sp- key, and the dedicated Token Plan endpoint.
2
Does the Output Limit Include Thinking?
Yes. The 393,216-token maximum, and the request-level Chat Completions output limit, cover thinking and the final answer together.
3
What Replaces the Preview After November 25?
QwenCloud’s cited release documentation does not name a replacement or general-availability date. Plan to return to an available model instead of assuming an automatic upgrade.
