Preview limited to Token Plan
China's Alibaba Group on Nov. 11 released a preview of "Qwen4-Max," the flagship of its next-generation AI model "Qwen 4." The official X account of the "Qwen" AI models announced it at 11 a.m. Nov. 11 Japan time. The model ID is "qwen4-max-preview," and it can also be called under the alias "qwen4.0-max-preview." It is available only to subscribers of "Token Plan," a flat-rate plan for developers. The company describes Qwen4-Max as a model with stronger knowledge, world understanding, programming and agent capabilities, in which the AI carries out work on its own, and says it is suited to complex programming, building web front ends and professional software work. The announcement said it would also be available early in "QwenWork," the AI coding tool "Qoder," the "Qwen" app and "QwenStudio."
Qwen4 is coming. Deeper knowledge. Richer world understanding. Stronger coding and agentic capabilities. Built for real-world tasks — from complex coding and knowledge work to front-end builds and professional software workflows. Qwen4-Max is now in preview on Alibaba's Token Plan.
Available until early Nov. 26, then withdrawn
The preview runs for about 15 days, from 10 a.m. Nov. 11 to midnight Nov. 25 China time (11 a.m. Nov. 11 to 1 a.m. Nov. 26 Japan time). Alibaba says it will not change the model during the period and will stop offering it when the period ends. Payment is made within the monthly flat-rate plan, and pay-as-you-go billing per token based on usage is not supported. Under Token Plan, each plan receives a set amount of "Credits" every month, which are consumed each time a model is used. During the preview, Qwen4-Max consumes as little as 20% of the usual Credits, falling a further 60% to 8% from 10 p.m. to 8 a.m. (China time). Alibaba Cloud's documentation for China lists the cheapest individual plan, Lite, at 39 yuan a month (20,000 Credits) and the higher-tier Pro at 312 yuan a month (200,000 Credits). The international version of Token Plan can be purchased and used only in the Singapore region, and is limited to interactive use by people in supported tools.
Reads 1 million tokens, accepts image and video input
According to QwenCloud's documentation, Qwen4-Max can read 1 million tokens at once. It can output up to 393,216 tokens per request, counting both the reasoning process before the answer (thinking) and the final answer. It accepts text, image and video input and outputs text. Thinking is enabled by default, and the depth of thinking can be chosen from three levels: "low," "medium" and "xhigh." The default is the deepest, xhigh. It supports Function Calling, which calls external functions, as well as web search and built-in tools, and can format answers as JSON. It is compatible with the OpenAI and Anthropic API formats and can be called from other companies' development tools, such as Anthropic's Claude Code. However, in the configuration example Alibaba Cloud gives for its AI coding tool "Qwen Code," the default output limit is 32,768 tokens, and users must set the limit themselves to get longer output. Alibaba's documentation does not say the model accepts audio or documents as input.
No benchmarks or weights released
Alibaba has not released benchmark results for the preview. It has also not released the parameter count, which indicates the model's size, a model card describing how it works, or weights that anyone can run. OrcaRouter, which runs an AI model routing service, reported that as of Nov. 11 the AI analysis site Artificial Analysis had no entry for Qwen 4. The company released the weights of some models in the previous generation, "Qwen3.8," and has not said whether it will release the weights of Qwen 4. BigGo Finance reported that performance figures and the release date of the official version have not been disclosed.
Name matched a leak that morning; timing later than expected
The name Qwen4-Max had come to light about seven hours before the announcement. LuminaBench, an X account that tracks AI model developments, posted at 4:22 a.m. Nov. 11 Japan time that "Qwen4 Max Preview" had appeared in the code of Alibaba's chat service "Qwen Chat." In the new build, the default model used by the agent feature had switched from the current "Qwen3.8 Max," it said. The post also said that the previous week had ended in disappointment after waiting for a release. Alibaba's documentation, however, does not confirm whether the preview is offered in Qwen Chat.
Qwen4 Max Preview is showing up in Qwen chat's code. The new build switches the agent mode default reference from Qwen3.8 max to Qwen4 max preview. Looking forward to this finally coming out after we got blueballed last week
Settings added to Qwen Code's public repository at the end of October anticipated the 1-million-token context and the three levels of thinking depth. On Oct. 29, event attendees reported that the chief technology officer (CTO) of the company's cloud unit had said "next week," but the actual release came about two weeks later. Information that predicted a release on Oct. 24 or within October proved wrong.
- Alibaba announces Qwen 4 is in trainingOfficial
- Information that Qwen4 Flash will be released Oct. 24 (wrong)Leak
- Rumor of "release within October" (wrong)Rumor
- 1-million-token context setting in Qwen CodeLeak
- CTO says "next week" (attendee reports)Report
- Three-level thinking depth setting in Qwen CodeLeak
- Official X account posts an image seen as a teaserOfficial
- "Qwen4 Max Preview" in Qwen Chat's codeLeak
- Qwen4-Max preview releasedOfficial
Earnings due Nov. 19
Alibaba is scheduled to report earnings for the July-September quarter on Nov. 19. The company has been expanding investment in generative AI and the cloud, and capital expenditure in the April-June quarter rose 75% to 67.7 billion yuan from 38.7 billion yuan a year earlier. Revenue from its AI cloud and computing services unit rose 45% to 48.4 billion yuan from 33.4 billion yuan a year earlier. Its U.S.-listed depositary shares closed down 2.52% from the previous day at $102.72 on Nov. 10.
