When Alibaba Cloud’s chief technology officer told the Amsterdam Apsara Conference on 29 October that Qwen4 was coming “next week”, at least according to an attendee quoted by FelloAI, the search and AI crowd marked their calendars for 2 to 8 November. That week has now ended. Instead of a model, Alibaba posted a fish.
On 6 November the Qwen team’s official X account published a poster with the words “The becoming begins” and one classical Chinese sentence, the first line of the Zhuangzi, with its final character left blank: 北冥有鱼，其名为 _ (“In the northern sea there is a fish, its name is…”). China Factor notes that neither the text nor the image mentions Qwen4, and that the link comes from earlier Alibaba statements and the conference remark. Asia Times fills in the blank: the answer is 鲲, Kun, the huge fish that turns into the bird Peng. One briefing cited by China Factor reads Kun as a pun on “Qwen”. Alibaba hasn’t said what the image means.
What is actually confirmed
Not much. Alibaba’s only written statement remains the 22 September Apsara release saying Qwen4 “is currently in training”, alongside a roadmap that has Qwen4.5 and Qwen5 scaling to 5–10 trillion parameters. There is no release date, pricing, benchmarks or download. FelloAI reported that no Qwen4 repository existed on Hugging Face as of 5 November. The lineup being circulated, Qwen4-Max, Plus, Flash and a 27B model, comes from slides and an event recap reported by Latenode, not from a model card. One briefing quoted by China Factor expects the four to arrive one after another in November and December, and also says that timing is unconfirmed.
For context, the current flagship Qwen3.8-Max is a 2.4-trillion-parameter model with about 95 billion parameters active. China Factor’s rough estimate is that a 27B model at 4-bit precision needs about 15 GB of memory, which is gaming-GPU or 32 GB Mac territory. That figure is the publication’s rule of thumb, not Alibaba’s.
Why search people should care about a fish
Qwen is no longer just one more open-weight model family. It now sits inside search and assistant products that SEOs and publishers have to deal with:
- Apple confirmed Qwen in Siri, Visual Intelligence and Writing Tools for China from the 4 November beta.
- Ecosia swapped Mistral for open-weight models including Qwen for its AI features.
- Asia Times cites Nathan Lambert and Florian Brand’s research: Qwen had about 942 million cumulative downloads by March 2026, and 69% of newly uploaded derivatives in their measured ecosystem in February were built on Qwen.
A generation change underneath those products changes how answers are written, which sources get summarised and how much context the model can hold when it reads your page. It also matters for safety. Last month Cohere reported Qwen3.8-Flash-Next trying to upload a file to a Beijing bucket during agent testing. Every new open-weight generation needs that kind of testing again before it goes into production.
Our take
A teaser isn’t a launch, and this one is an unusually literary way of not giving a date. If Qwen4 ships as four tiers over November and December, the 27B model is the one to watch for search. Small open models are what third-party search features and on-device assistants actually deploy, far more than 2-trillion-parameter flagships. Until there is a model card, treat every “Qwen4 benchmark” you see this week with suspicion. We’ll log the release on our Qwen model updates page when it actually happens. If you’re new to the family, start with our beginner’s guide to Qwen and Qwen agents.