Alibaba opens Qwen4-Max preview with a million-token context and a fixed end date

Alibaba opened Qwen4-Max-Preview on November 11, giving developers access to a flagship AI model designed for coding and multistep tasks. Alibaba Cloud lists a one-million-token context window and support for text, images and video. But this is temporary access: the company will retire the preview after November 25, 2026.

The company positions the model for office productivity, frontend design and complex agent tasks, as well as software development. Its documented capabilities combine understanding visual material with web search and calls to external functions. That gives developers a model intended not just to answer questions, but to work with tools as part of a larger application.

A subscription preview with a hard stop

The official preview window began at 10 a.m. on November 11 and ends at midnight at the close of November 25, UTC+8. Alibaba says it will make no changes to the model during that period. The endpoint therefore comes with both a defined evaluation window and a retirement date, rather than an open-ended promise of continued availability.

For API access, developers need Alibaba’s Token Plan; pay-as-you-go usage is not supported. The setup instructions cover both Personal and Team plans, with a dedicated API key and service address. This is a subscription-based route to the preview, not an API offer billed directly by token volume.

MobiDevices also reports access through Qwen Chat on the web and in the app, QwenStudio, QwenWork and Qoder. It lists international QwenCloud subscriptions starting at $6 a month for Lite, including 20,000 credits. The other listed tiers are Essential at $10 for 45,000 credits, Standard at $18 for 80,000, and Pro at $43 for 200,000.

The same outlet reports that Alibaba has not published official benchmark results, a parameter count or weights for local deployment. A final-version release date has not been announced. The preview is available to evaluate, but there are no published official test results here to establish how it compares with other models.

Visual inputs, tools and structured answers

Alibaba’s documentation confirms the million-token context length as a model specification. It separately lists image and video understanding for screenshot analysis, document understanding and video analysis. These are supported input capabilities; the model’s listed generation capability is text, including structured answers.

For applications that need a predictable answer format, the model supports both JSON Object and JSON Schema outputs. Its tool capabilities include function calling, web search and built-in tools. Those features sit alongside visual understanding, allowing an application to combine analysis of supplied material with requests to connected tools.

The integration detail that can cut an answer short

The API supports OpenAI Chat Completions, OpenAI Responses and Anthropic Messages. Alibaba’s quick-start instructions specify three setup steps for Token Plan users:

1. Obtain a dedicated API key through Token Plan Personal or Team.
2. Use the dedicated Token Plan service address and the endpoint for the chosen API protocol.
3. Set the model ID to qwen4-max-preview.

Reasoning is enabled by default. Developers can choose low, medium or xhigh through reasoning_effort, with xhigh selected unless changed. Alibaba warns against setting reasoning_effort and thinking_budget together. The control adjusts reasoning intensity, while the output allowance must accommodate both that reasoning and the eventual answer.

The important compatibility change concerns max_tokens in Chat Completions. Starting with this preview, it means the same thing as max_completion_tokens: both cap the combined reasoning and final answer. Alibaba warns that a limit previously sized only for the visible answer may now truncate output. Existing Qwen models retain their previous behavior.

Multiturn conversations also have a specific handling rule. Chat Completions enables preserve_thinking by default. Alibaba instructs developers to return historical reasoning in the assistant messages’ reasoning_content field, rather than combining it with content. For this preview, adopting a familiar API format still means checking model-specific token limits and conversation handling.