The week people have been counting is over, and the model is not in it. On October 29, at Alibaba Cloud's Apsara conference in Amsterdam, attendees reported that CTO Feifei Li said "Qwen4 next week." Read as a calendar, that is November 2 through November 8. Alibaba's own write-up of the Amsterdam event does not mention Qwen 4. On Sunday evening, November 8, a search of the official Qwen organization on Hugging Face for Qwen4 returned no model. The cards at the top of that page are still Qwen 3.8.

What Alibaba has put in writing is older, and narrower. The September 22 press release from the Hangzhou Apsara conference says Qwen 4 "is currently in training." It gives no day, no sizes, and no license. The same release puts 5 to 10 trillion parameters on Qwen 4.5 and Qwen 5. That scale is the roadmap past Qwen 4. It is not a Qwen 4 size.

The four names people repeat come from a stage slide, not from that press release. Coverage of the Amsterdam talk describes a slide, "Qwen4 Series Coming Soon," listing Qwen4-Max, Qwen4-Plus, Qwen4-Flash, and Qwen4-27B. No parameter count, no price, no context length, and no license sits under those names. A 27B on a slide means a size they are willing to show. It does not mean a file you can download, and it does not mean the Apache 2.0 license that the last 27B used.

The new official post does not close any of that. On November 6 at 12:00 GMT the Qwen account published four words, "The becoming begins," with a poster. The poster has the Qwen mark, a flock of birds over water, a classical line whose name is left blank, and the same English sentence. It does not say Qwen 4. When we opened the post on November 8 it was past 2.5 million views and past 18,000 likes. Unsloth, the group that ships GGUF files for local runners, replied two hours later: "We're so ready! We can't wait to upload GGUFs for people to run Qwen4 locally." That is a converter saying the tools are waiting. It is not a weight, and it is not a date.

The code is further along than the poster, and it is still not a model card. On October 28 QwenLM merged a change to Qwen Code, pull request 15762, titled "fix(core): size the qwen4 generation at 1M." Any id that begins with qwen4 is now given a 1,000,000-token context in that command-line tool, ahead of the old 262,144 fallback. The names the tests cover are qwen4, qwen4-max, qwen4-plus, qwen4.5-max, qwen4-coder-plus, a dated spelling qwen4-max-2026-10-01, a provider-prefixed openai/qwen4-max, and qwen4-omni-120b-think. The dated spelling is a pattern the detector has to recognize. It is not evidence that a model shipped on October 1. The 120B omni name is a test fixture already in the repo. The author says it is not a released model. The change was checked against the repository's own fake server. No real Qwen 4 endpoint was called. The pull request says the 1,000,000 figure came from a reviewer, and that neither the code nor the models.dev catalog can confirm it. The output ceiling was left at the family's 32,768-token default. A catalog entry, once one exists, overrides the guess.

Several numbers now circulating are about other models. Qwen3.8-Flash-Next is a model you can already download. Its card is the 125 billion model with 6 billion active and a separate 51 billion n-gram table. The Qwen organization page currently lists that repository at 180 billion. An architecture tag on that card, qwen4_exp, is a preview label. It does not rename the checkpoint Qwen 4, and it does not publish a Qwen4-Flash parameter count. Qwen3.8-27B is the local model people already run. The repository was created on August 5, 2026, and the card carries an Apache 2.0 license. The open weights of the 2.4 trillion Qwen3.8-Max, published as Qwen3.8-2.4T-A95B, have a repository created on August 8. Prices quoted for Qwen3.8-Flash and Qwen3.8-Max are prices for those models. No Qwen 4 price has been posted.

Two later dates are being treated as if they were announcements. APIMaster says community reports point at Tuesday, November 10, and says the original public source was not supplied. Modelcast's own forecast, updated November 6 and checked again November 8, puts the most likely day at November 18, inside a window that runs to January 15, 2027, and calls that a 63 percent chance. Both are forecasts. Neither is a sentence from Alibaba.

The desk's call is a week, not a Tuesday. Friday's poster is how a launch week starts when the names are already compiled into the tool and the weights are not up. We put the first public Qwen 4 name, an API id or a repository that finally uses the name, in November 10–14. If that Friday is also empty, November 18 is the next dated guess on the board, and a miss there is a December. The 27B is the part we do not push a month out. In the 3.8 generation the open 27B repository appeared on August 5, before the 2.4 trillion weight repository on August 8. The local model was not the one they held back. A Qwen 4 27B file, if they repeat that order, comes with the first named release or within a few days. The license is the part we will not guess out loud. The last 27B was Apache 2.0. The Max weights were a different license. A slide that says 27B does not choose between those.

A LocalLLM thread making the rounds under the title "qwopeful" is the same mood as Unsloth's reply. Reddit blocked a clean read of that post from here. The readiness you can check without the thread is the poster, the reply, and pull request 15762.

If a Qwen4 repository or a Model Studio id appears, the week that failed becomes a delay. Until one of those exists, the poster is a blank where the name goes, and the million-token window is a client getting ready for a server that has not answered.

Our day is not November 10. Nobody has shown where that Tuesday came from. The desk's window for a first public name is November 10–14. The 27B download, if Qwen 4 follows Qwen 3.8, comes with that name or a few days after it. November 18 is the fallback if this week is another poster.