Qwen4-Max Opens in Preview With a 1M-Token Context

Alibaba has opened a preview of Qwen4-Max, the flagship of its next-generation Qwen family. According to the QwenCloud developer documentation, the model, with the API name qwen4-max-preview, went live at 10:00 Beijing time (UTC+8) on November 11. Qwen’s announcement on X opens with “Qwen4 is coming” and promises stronger coding and agentic abilities for real-world work. Besides the API, TechNode reports early access in the Qwen app, Qoder, QwenWork and QwenStudio.

The documentation lists a 1-million-token context window and a maximum output of 393,216 tokens, a budget that thinking and the final answer share. The model takes text, images and video as input, which Alibaba suggests for screenshot analysis, document understanding and video analysis. Thinking is on by default. The reasoning_effort setting accepts low, medium or xhigh, and xhigh is the default; it should not be combined with thinking_budget. The model supports function calling, web search and built-in tools, JSON Object and JSON Schema outputs, and speaks the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages protocols.

One change can catch existing integrations out. For this model, max_tokens in the Chat Completions API means the same as max_completion_tokens: both cap thinking and the answer together, so a limit sized for the answer alone can leave the reply cut short. Alibaba says other Qwen models keep their old behaviour. In multi-turn chats, preserve_thinking is on by default, and earlier reasoning goes back in each assistant message’s reasoning_content field rather than in content.

Token Plan only, for two weeks

There is no pay-as-you-go access. The preview runs only through the Token Plan, QwenCloud’s subscription billed in Credits, which comes in an Individual edition (Lite, Essential, Standard and Pro tiers) and a Team edition (Standard, Pro and Max seats). The plan is sold and served only in the Singapore region, through its own endpoint and its own API keys, and is built to plug into coding tools such as Claude Code, Cursor, Qwen Code, Codex, Qoder and OpenClaw. The overview page lists no prices.

The window is short and fixed. The documentation says the preview runs until midnight (UTC+8) at the end of November 25, that the model will not change during that period, and that it will be retired once the preview ends. Alibaba has not said what replaces it, or when a full Qwen4-Max release follows.

What the preview lacks is as telling as what it offers. Alibaba has published no benchmark table, no model card, no parameter count, no license and no weights. The 1-million-token context is also not new for the line: Qwen3.7-Max reached it in May, and in August Alibaba took Qwen3.8-Max global and promised open weights to follow. A frozen, two-week, subscription-only window reads less like a launch than a public trial, and the numbers that would place Qwen4 against its rivals are still to come.