Alibaba opened Qwen4-Max-Preview on November 11, 2026, giving Token Plan subscribers access to a flagship multimodal model with a one-million-token context. The release supports reasoning, image and video understanding, tool use, and several familiar API formats. It is also temporary: Alibaba says the model will be retired when the preview ends on November 25.
The access conditions matter as much as the specifications. Pay-as-you-go access is not supported, and developers need a dedicated Token Plan API key and endpoint. Alibaba’s model documentation also changes the meaning of an existing output-limit parameter, creating a potential source of truncated answers for applications carrying over older settings. These conditions are documented on the official Qwen4-Max-Preview page.
For developers, I’d treat this as a scheduled evaluation opportunity. The published retirement date makes it a poor dependency for anything expected to keep working unchanged after the preview.
What Alibaba Has Opened, and for How Long
Alibaba’s Token Plan release log dates the launch to November 11. The model documentation gives the precise access window:
- Starts: November 11, 2026, at 10:00 UTC+8.
- Ends: November 25, 2026, at 24:00 UTC+8, meaning midnight at the end of that day.
- UTC equivalent: November 11 at 02:00 through November 25 at 16:00.
- Afterward: The preview model will be retired.
Alibaba also says no changes will be made during this period. That gives developers a defined evaluation window, but the cited announcement provides no promise of a replacement model or continuing access under the same ID.
Alibaba positions Qwen4-Max-Preview for coding, office productivity, frontend design, and complex agentic tasks. Its published specifications describe the following capabilities:
| Item | Documented support |
|---|---|
| Model ID | qwen4-max-preview |
| Context length | 1 million tokens |
| Maximum output | 393,216 tokens, shared between reasoning and the final answer |
| Input types | Text, images, and video |
| Reasoning | Enabled by default |
| Reasoning intensity | low, medium, or xhigh, with xhigh the default |
| Tools | Function calling, web search, and built-in tools |
| Structured output | JSON Object and JSON Schema |
| API compatibility | OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages |
The one-million-token context is a capacity specification, not a one-million-token answer allowance. Alibaba separately caps output at 393,216 tokens, and that output allocation includes reasoning.
A larger context can make longer conversations, document collections, or codebases possible within a request. It does not by itself tell developers how reliably the model retrieves a small detail from that material. The cited model pages provide capabilities and limits, but no benchmark results or independent measurements establishing reasoning quality, long-context accuracy, latency, or coding performance.
My view is that the useful evaluation question is whether the model completes a particular workflow accurately within the available time and budget. The context headline alone cannot answer that.
Token Plan Access Requires More Than a Model-Name Change
Alibaba’s quick-start instructions require a Token Plan subscription, a dedicated API key, and the correct service endpoint. Changing an existing request’s model field to qwen4-max-preview is only one part of the setup.
The two official documentation sites list different base URLs:
| Documentation environment | Published Token Plan base URL |
|---|---|
| QwenCloud | https://token-plan.maas.qwencloudapi.com |
| Alibaba Cloud Model Studio | https://token-plan.ap-southeast-1.maas.aliyuncs.com |
The QwenCloud guide names Token Plan Individual and Team, while the Alibaba Cloud guide names Personal and Team. Each directs developers to its Token Plan instructions for the complete endpoint associated with their chosen API protocol.
The practical rule is to keep the subscription, credentials, base URL, and protocol instructions aligned with the service being used. Neither page presents its base URL as a complete request path for every supported API. Mixing setup details from the two environments is an avoidable troubleshooting risk.
API Compatibility Does Not Remove Plan Restrictions
S5 Labs reports that both Token Plan Personal and Team restrict access to interactive use within coding and agent tools. Its reading of the plan documentation says automated scripts, custom application backends, and non-interactive batch calls are prohibited.
That account is secondary reporting on plan terms; the model page itself does not spell out those restrictions. Developers should check their applicable Token Plan agreement before connecting an automated evaluation harness or application backend.
This distinction is consequential. Compatibility with OpenAI or Anthropic request formats describes how software can communicate with the model. It does not grant permission to use every workflow those APIs could technically support.
I would resolve that permission question before building a benchmark runner. A documented API format and an active subscription are insufficient grounds to assume unattended testing is allowed.
The Output-Budget Change Can Truncate Answers
The most consequential migration detail concerns OpenAI Chat Completions.
Starting with qwen4-max-preview, Alibaba says max_tokens is equivalent to max_completion_tokens. Both parameters limit the combined tokens used for reasoning and the final answer. Existing Qwen models retain their previous behavior, according to the official usage notes.
That creates a subtle configuration risk: an old parameter can remain valid while its effect changes.
Suppose a request has an 8,192-token completion limit. If reasoning consumes 6,000 tokens, at most 2,192 tokens remain for the final answer. This is an illustrative calculation, not a measured Qwen4 result. It shows why a limit previously chosen around the expected answer length may now be too small.
Thinking is enabled by default, with xhigh as the default reasoning intensity. Developers should therefore review both the completion allowance and the reasoning setting.
An illustrative Chat Completions request body could look like this:
{
"model": "qwen4-max-preview",
"messages": [
{
"role": "user",
"content": "Review this proposed database migration and list the main risks."
}
],
"reasoning_effort": "medium",
"max_tokens": 16384
}
This example shows the relevant fields; it is not a complete client setup or a reported test. The 16,384-token limit is shared by reasoning and the visible response. It is an example allowance, not a guarantee that the answer will fit or a recommended size for every task.
Alibaba supports low, medium, and xhigh for reasoning_effort and explicitly warns against setting reasoning_effort and thinking_budget together.
I’d make output completeness a specific evaluation criterion when migrating an existing integration. A shorter response could reflect budget exhaustion, so comparing visible answers without checking configuration could lead to the wrong conclusion about model quality.
Frequently Asked Questions
3 questions
1Is Qwen4-Max-Preview Available Pay as You Go?
No. Alibaba lists the preview as Token Plan-only and requires a dedicated Token Plan API key.
