“Next week” turned out to mean two weeks. On 11 November Alibaba’s Qwen team posted that Qwen4-Max is now in preview, five days after the “The becoming begins” Zhuangzi teaser we wrote about yesterday. The post went up at 10:01 Beijing time (02:01 GMT), according to Pasquale Pillitteri’s write-up, which quotes it in full:
“Qwen4 is coming. Deeper knowledge. Richer world understanding. Stronger coding and agentic capabilities… Qwen4-Max is now in preview on Alibaba’s Token Plan. You can also get early access to the new model on QwenWork, Qoder, Qwen APP, and QwenStudio.”
Where you can get it
There are five routes in: QwenCloud’s Token Plan (the credit-based subscription for coding and agent tools), plus QwenWork, Qoder, the Qwen app and QwenStudio. TechNode, citing QwenCloud and Cailian Press, says the model supports image and video inputs and up to a million tokens of context. Pillitteri reports the Token Plan model ID as qwen4-max-preview, with Lite plans from $3 a month for new users, and says the Token Plan preview closes on 25 November Beijing time, citing a Japanese guide to the plan terms.
The clues were already in the code. ArtificialWatch reported that Qwen Chat front-end build 0.3.40, live at 19:52 UTC on 10 November, set qwen4.0-max-preview as the default model for agent mode, replacing qwen3.8-max, and added an agent_qwen4 announcement pop-up. At that point the public model list still showed Qwen3.8-Max.
What we still don’t know
Nvdium is blunt about it: the official post has no model card, no parameter count, no context window, no API model ID and no official benchmarks. “What is confirmed is where to try it, not what the model is made of.” The million-token figure appears in TechNode’s report and in Qwen Code client settings, according to Pillitteri, but not in the launch post. There are no open weights and no licence. Plus, Flash and the 27B tier from the roadmap are not part of today’s preview. Discount figures disagree between sources too (80% on the plan page versus “up to 90%” elsewhere), so check at checkout.
Why this matters for search
Qwen is more than a chatbot story for us. Alibaba’s models sit behind the Qwen app and its agents and, as Apple confirmed this month, behind Siri, Visual Intelligence and Writing Tools in China. A new flagship tends to filter into all of those answer surfaces over the following months. The emphasis on “agentic capabilities” and “real-world tasks” also matches what every Chinese search player is pushing this autumn: search as something that completes a task, not just a page of results.
Our take
It’s a preview with a deadline and no spec sheet. Treat it as a chance to test, not something to cite as fact. If you track how Chinese AI assistants cite and summarise your content, run the same prompts through Qwen3.8-Max and Qwen4-Max-Preview now, while both are on the Token Plan, and log which sources each one pulls. That baseline will be useful once the model reaches the consumer Qwen app by default. We’ll add Qwen4-Max to our Qwen model updates page once Alibaba publishes a model card, and we won’t put guessed figures in the table before then.