The only date in Alibaba's writing is no date. The September 22 press release says Qwen 4 is currently in training, and it puts 5 to 10 trillion parameters on Qwen 4.5 and Qwen 5. Coverage of the Hangzhou keynote adds "very soon." A CSC Financial note on September 28 called the release imminent. Neither line names a day.

The numbered guesses before this week are already dead. A prediction market in early September priced a launch before October 1 at 44 percent, and October opened without one. By October 26 a market of the same kind sat at 83 percent for a launch before November 1, as Yotta Labs recorded it that day. November opened with no Qwen 4 weights. Countdown pages that had circled October 30 rewrote themselves once that day passed with no repository.

What replaced October is a sentence from a room. On October 29, at Apsara in Amsterdam, attendee John T Davies quoted Alibaba Cloud CTO Feifei Li, not the Stanford researcher of the same name, saying "Qwen4 next week." The slide in the room still read "Qwen4 Series Coming Soon." Next week from October 29 is November 2 through November 8. Today is the first of those days. On October 31 LuminaBench wrote that its expectation is the week beginning November 2, and that the Qwen Code change it was reading does not contain a launch date. That is the rumor with a person attached to it. It is not a press release.

The other live rumor is the one the calendars wrote after October failed. A countdown re-read on November 1 sets Qwen 4 Max at November 16 and marks the estimate low confidence. The page's method is the gap between Qwen Max updates, May 19, August 3, and September 2, then two more weeks past the missed late-October leak. Modelcast, updated November 2, calls November 18 its most likely day. It says the pick rests on four signals, and that one of the four is its own estimate. The window it draws around that day runs from November 2 out to January 15, wide enough to survive being wrong. A roundup posted the same day by Mark Kretschmann puts Qwen 4 Max at about November 16 and says in the post that none of the dates on his list are confirmed.

Our read of that pile: the rumor that can still be on time is this week. If November 8 ends with no model card, the date with a method behind it is November 16, and November 18 is the same bet. A miss there does not produce a new day. It leaves Modelcast's outer edge, January 15, and that is no longer a date.

The names hung on whichever day wins are the four from the stage: Max, Plus, Flash, and a 27B for local machines. An Intel kernel merged October 23 describes that 27B as dense, with the same 5,120-wide hidden state as Qwen3.8-27B. That shape is a reading of kernel code, not a config. The feed-forward width inferred from the same code is not a published size, so it stays out. The architecture that activates 6 billion weights out of 125 billion already shipped on August 26 as Qwen3.8-Flash-Next. A 1,000,000-token figure in Qwen Code is a client fallback for ids that are not in the catalog. None of that dates the release. It is what people expect to find if this week, or the 16th, actually hits.

The hole is a repository. Until one exists, both dates are rumors, and the earlier one expires on Sunday, November 8.

If November 8 passes empty, the rumor that still has a method is November 16. November 18 is that guess plus two days.