Waiting for Qwen4: what Alibaba actually said on 22 September

Since late morning a graphic has been moving through the feeds: four tiles labelled Qwen4-Max, Qwen4-Flash & Qwen4-Plus, Qwen4-27B, with “Qwen4 series coming soon” above them. Attached is the claim that Alibaba announced exactly these models at the Apsara Conference in Hangzhou, and that several of them can run on local hardware. We approach that the way we approach any model that has allegedly just shipped: measure it, don't retell it. As of 22 September, 11:00 CEST, there is no Qwen4 model you can hold.

What is evidenced and what is not

The story has a hard core and a soft crust. The core is verbatim in Alibaba's press release of 22 September: Qwen 4 is “currently in training”, plus a roadmap for Qwen4.5 and Qwen5 at 5 to 10 trillion parameters. Alongside it, the products actually announced: the Zhenwu V900 accelerator, Qwen3.8-LiveTranslate, Qwen-Audio-3.1 in three variants, Qwen-Image-3.1 later this year, and the Qwen Intelligence platform. The trade press sitting in that hall on the day (Fast Technology via NetEase, IT之家) reports precisely those points and no model family.

Qwen is Alibaba Cloud's model family. The open line runs through Qwen3.6 and Qwen3.7 to Qwen3.8; so far, Qwen4 exists only as a name in code. (Quelle: Wikimedia Commons, file Qwen_Logo.svg)

Claim	Checked on 22 September	Result
Four Qwen4 models were announced	Alibaba press release	none of the four names appears
The models are already orderable	Model Studio catalog, state of 22 September	qwen3.8-max, qwen3.7-plus, qwen3.8-flash
The weights are downloadable	Hugging Face API, queried by us	468 models in the Qwen org, none named Qwen4
Runnable locally	our own calculation	nothing that has a file

Where the graphic comes from

The four names surface in three places, and none of them is Alibaba. An X post from the venue lists them verbatim (@CuiMao, “Qwen4 full family”), a small navigation aggregator describes a slide marking Qwen4 as coming soon (chooseai.net, 83 views, its lead summary machine-generated), and a thread on linux.do repeats the same list. That is one report passed through three mouthpieces, not corroboration.

The clearer tell sits in the margin. The same sentence about “four core models” also appears as a boilerplate block on automatically generated Weibo pages about a film release, a used-car dealer and celebrity gossip; SEO filler pastes that sentence onto any context. And the graphic is not a photo of a slide: it reproduces the X-post enumeration one to one, including the category error of placing a size step, Qwen4-27B, beside the tariff tiers Max, Plus and Flash. We queried the hub's API ourselves with the search term Qwen4: zero hits under the Qwen organisation, in return third-party test fixtures that carry the architecture name in their repository name.

The architecture is already here, under a different name

Correcting the story turns out to be more interesting than the story. What is meant by the Qwen4 architecture has been sitting openly on the hub since 26 August: Qwen3.8-Flash-Next, labelled on its own model card an experimental preview of the architecture that will underpin Qwen4. We read config.json ourselves rather than the blog post: architectures = Qwen4ExpForConditionalGeneration, model_type = qwen4_exp, 512 experts with 10 routed per token plus one shared, 48 layers of which 36 use linear attention via Gated DeltaNet and 12 use Qwen Sparse Attention, plus gated residual and the n-gram table. 125 billion parameters in total, 6 active per token.

That also settles the question asked directly underneath the graphic, whether this is the end of the Mixture-of-Experts. It is not. The model stays MoE and adds a second axis next to it: n-gram embeddings with 51 billion entries, justified by Alibaba with the argument that such a table offloads more easily than extra experts. Anyone reading Qwen4 as a departure from MoE has not read the code, which already exists.

What of this fits on our cards

Our inference host carries four RTX 5090s with 128 GB of VRAM in total, documented in our own AI stack. For Qwen4 there is no calculation to make on that box, because there are no weights. The calculation that does exist is the architecture's n-gram table: it ships as a separate 47.7 GiB file in the hub, which makes it larger than our active share of the main weights. The open pull request sgl-project/sglang#40235 from 18 September tries to back that table with host memory through a file-backed store, and reports, for FP8 at tensor rank 4, 71 GiB of pinned host RAM without a cap, 49 GiB with a 64 GiB cap, 15 GiB at 32 GiB, 6 GiB at 24 GiB, at the price of 9.2 instead of 8.6 ms per token.

The largest step sits at the first cap: 71 down to 49 GiB, with latency up six percent. (Quelle: Our own reading of the open SGLang pull request 40235; figures are the author's)

Two more items that caught us twice in 2026. First the licence: Qwen3.8-27B is Apache 2.0, Flash-Next carries the Qwen Community Model License 1.0 and Qwen3.8-2.4T-A95B a bespoke Qwen3.8-Max licence, so twice not Apache. Before any commercial use that means opening the LICENSE file in the repository instead of reading the banner on the hub. Second, the memory shuffling: in an architecture where tables, KV cache and active share compete for the same silicon, the problem migrates to wherever nobody is looking. For us that was the KV cache, in the four-day test and in the follow-up through rented cloud.

What we are waiting for

We are not waiting for the announcement, we are waiting for the repository. Three observations suffice, and we need nobody else's opinion on any of them:

A repository on the hub. A model named Qwen4 under the Qwen organisation turns the family into fact. Until then it is a rumour with good source code.
An ID in the catalog. qwen4-max in the Model Studio model overview would be the same signal from the selling side. That catalog carries a 22 September date and still says Qwen3.8.
A published slide. qwen.ai or the Qwen blog with the family picture, and we can talk about tariffs, context lengths and licences instead of assumptions.

The conference runs to 24 September, so something may still land. The bar on our side is already set: whatever arrives in the 27B segment has to get past Qwen3.8-27B in self-hosted operation, the first model in our size class that delivers reliably for us. If Qwen4 turns out as good as 3.8, that is a good result and not a given. Until then we run the same route as always: isolated environment, our task set from the self-hosting post, comparison against the lane that is already running, and only then a word about deployment.