What Amsterdam actually confirmed
The conference ran at Beurs van Berlage, Amsterdam's old stock exchange, the first Apsara edition in Europe after years in Hangzhou. Feifei Li opened the day; at Alibaba Cloud he is both CTO and president of the international business.
Two announcements were official. Alibaba Cloud opened its first cloud region in the Netherlands, with two data centers, its fourth in Europe after Frankfurt, London and Paris. The same day Qwen3.8-Flash became available in the EU through Model Studio.
Qwen4 shows up in none of those texts: the day's story rests on a sentence spoken in the room and on a slide.
Where the line Qwen4 next week came from
The date did not come from Alibaba. An attendee wrote on X that CTO Feifei Li said "Qwen4 next week" from the stage. Minutes later the same person posted a photo of a slide from Qwen researcher Xudong Guo: "Qwen4 Series Coming Soon", with four names, Qwen4-Max, Qwen4-Flash, Qwen4-Plus and Qwen4-27B.
Read literally, "next week" means November 2 to 8. It is easy to treat that as a commitment, and some people already are. I would not. The sentence was spoken in a hall, one person carried it out, and no Alibaba or Qwen channel has repeated it.
Why the Coming Soon slide outweighs the quote
"Coming soon" is the phrase companies use when they will not put a date in writing. A projected slide sits closer to the official position than a paraphrase from the audience, and the slide says soon.
The Qwen team also ships models without warning: a post on X, weights on Hugging Face, done. That habit makes every announced date soft.
My bet: the first weights land in November, likely in the first half of the month. I will not name a day.
The clue in the code: a million tokens in Qwen Code
On October 28 the Qwen Code repository merged change 15762. It assigns the whole qwen4 family a context window of 1,000,000 tokens. Before that, qwen4 identifiers fell back to a default of 262,144.
The author of the change is explicit: a reviewer passed the number along, he did not verify it himself, and the tests ran against a mock server. So the million tokens live in a tool's configuration, and the model spec is silent.
The code mentions qwen4-max, qwen4-plus, qwen4-coder-plus, qwen4.5-max and qwen4-omni-120b-think. The last one hints at a mid-sized multimodal model with reasoning switched on, though that is a reading of a name.
What Alibaba said in September
On September 22, at Apsara in Hangzhou, the company said on the record that Qwen 4 was in training and showed four names: Max, Flash, Plus and 27B. The same keynote announced Qwen 4.5 and Qwen 5 at 5 to 10 trillion parameters. The flagship Qwen3.8-Max sits at 2.4 trillion, as Reuters reported.
No license for Qwen4 has been announced. In the 3.8 generation Alibaba published Flash and 27B with open weights, and served Max and Plus mostly through its API. If the pattern holds, small studios will run the 27B locally.
Table: what is confirmed, what is rumor
A slide and a spoken sentence carry different weight.
| Claim | Status |
|---|---|
| The "Qwen4 Series Coming Soon" slide and four names | Confirmed by a photo from the room |
| The "Qwen4 next week" line | Attendee's account on X, never repeated officially |
| A release date of November 2 to 8 | Confirmed by neither Alibaba nor the Qwen team |
| A context of 1,000,000 tokens | Code change in Qwen Code, value unverified by its author |
| Open weights for Qwen4 | Not announced, community expectation |
| The Dutch region and Qwen3.8-Flash in the EU | Official, October 29 |
Why this matters if you produce video
Qwen and the Wan video model belong to the same Alibaba family. Wan3.0 went generally available on August 24, and the pairing of a language model with a video model inside one cloud works as a single tool: script, shot list, per-shot prompts, then a check on the result. Any stack for AI video production deserves a look at the whole chain, since half the hours go into the text around the picture.
That is where a million tokens would matter. A twenty-minute episode script, a shot list, a prompt library and a revision log rarely fit inside 262,000 tokens together, while a million swallows the package and lets an agent hold a whole episode without losing continuity between shots.
This is a hypothesis built on someone else's repository config. Rebuilding a pipeline on it would be premature, yet testing is cheap: Qwen3.8-Flash is open and shows how the architecture handles long context.
My take: how to prepare for the release
Do not rewrite production around a rumor. A date spoken in a hall can slip by a week, and the work still costs money.
Test what is already open. Qwen3.8-Flash is the architectural preview of Qwen4, a mixture of experts with 125 billion parameters plus 51 billion N-gram embeddings. A comparison against the current flagship on your own tasks beats any forecast.
Keep the 27B in view, since it shipped open in the 3.8 generation. Read the license too: for commercial work the terms of use outweigh a couple of points on a benchmark.
Volume decides the rest. One video a month runs fine on a subscription, while vertical video ads and short-form video at scale need another setup: one template, many versions, a predictable cost per unit. Where a real person has to be on camera, handcrafted videos remain the stable part of the budget. For a step-by-step walkthrough, our guide on how to make AI video covers the current lineup.
