Alibaba Gives an Early Look at Qwen4: Qwen4-Max Preview Reads a Million Tokens at Once, but It Expires in 15 Days and Comes With No Benchmarks

Qwen, one of the most widely used open model families in the world, is getting a new generation.

On Nov. 11, Alibaba's Qwen team posted that "Qwen4 is coming," promising deeper knowledge, richer understanding of the world, and stronger coding and agentic abilities. At the same time, a preview of the flagship Qwen4-Max went live on Alibaba's Token Plan subscription service, and the company's own apps, including the Qwen app, the QwenWork office suite, the Qoder coding tool, and QwenStudio, got early access.

Why Qwen matters

Qwen is Alibaba's in-house family of large language models. Over the past two years, many of its versions were released with open weights, downloaded heavily, and used as the starting point for countless smaller and specialized models in China and abroad. When Qwen moves to a new generation, it affects not just Alibaba's products but a large community of developers who build on it.

What Alibaba actually shipped

According to Alibaba Cloud's documentation, Qwen4-Max-Preview is a multimodal flagship: it reads text, images, and video; it "thinks" before answering by default; and it can call tools, search the web, and return structured output in a format you specify.

Two numbers are worth unpacking. First, the context window is 1 million tokens. A token is roughly a word fragment, so this is enough to hold hundreds of thousands of words at once, such as the code for a mid-size software project or several books. Second, the model can produce up to 393,216 tokens in one reply, a budget shared between its internal reasoning and the final answer.

For developers, a practical detail: the model is compatible with OpenAI's and Anthropic's API formats. In plain terms, a coding tool set up for GPT or Claude can try Qwen4-Max by swapping in a different endpoint and key.

Cost and access

The preview is available only through the Token Plan subscription, not pay-as-you-go. Alibaba advertises discounts of up to 90% during the promotion. Chinese outlet IT Home reported that Token Plan's personal tier starts at 39 yuan (about $5.50) a month, with new users getting another 50% off during the event.

The fine print

This is a time-limited preview. Alibaba's documentation says it runs from 10:00 on Nov. 11 to midnight on Nov. 25, Beijing time, with no changes in between, after which the model will be retired. Alibaba hasn't said when the full Qwen4 lineup will arrive.

Alibaba also hasn't published benchmark scores, a parameter count, a technical report, or whether the weights will be released openly as with earlier Qwen models. "Stronger coding and agentic abilities" is, for now, the company's own claim, with no independent testing yet.

What comes next

Baidu launched its ERNIE 5.5 Flash Preview the same day, and several Chinese labs have released models in quick succession in recent weeks, all competing hardest on coding and on AI agents that carry out multi-step tasks on their own. Once Alibaba publishes full Qwen4 specs and test results, they will be a key yardstick for how close Chinese models are to the top U.S. systems.

What it means for you

If you use the Qwen app or its office tools, you may already be using the new model; watch whether long documents, spreadsheets, or writing tasks feel different. If you're a developer, you can try a new Chinese flagship for a few dollars, but don't build anything important on a preview that disappears on Nov. 25.

Original sources published: Qwen official X account (@Alibaba_Qwen), Nov. 11, 2026, 10:01 Beijing time (UTC+8); IT Home, Nov. 11, 2026, 10:10 Beijing time; Sina Tech, Nov. 11, 2026, 11:08 Beijing time