Qwen 4: what the line is, whether it is out, and what Alibaba has actually said

Qwen 4 is not available, and no release date has been announced for it. That is the whole answer to the two questions most people arrive with, and it is worth stating before anything else, because the Qwen 4 line — Qwen 4 Max, Qwen 4 Flash, Qwen 4 Plus and Qwen 4 27B — has been named in public by A​libaba but never shipped in any form: no model card, no weights, no API identifier, no licence, no context window, no price. The only line you can actually call or download today is the 3.x generation, whose current flagship is Qwen3.8-Max. This page is the standing reference for the name itself: what has been said on the record, what has only been inferred, and how to tell the two apart.

On the size of the question, since it is the reason this page exists rather than another announcement post: across the two spellings a search engine splits it into — qwen 4 and qwen4 — our own Search Console property held 36,895 impressions and 2,249 clicks over the year to 13 October 2026, and the largest qualified row inside that total is qwen 4 release date at 3,434 impressions. Six of the top eight rows are asking when. Nothing on this site answered the plain version of the question, so this is it.

The one sentence Alibaba has written about Qwen 4

The written record is short enough to quote. In its press release "Alibaba Unveils Roadmap on Full-Stack AI Strategy from Chips, Cloud Infrastructure, Models to Agents", dated 22 September 2026 and published out of the Apsara Conference in Hangzhou, Alibaba states that its next-generation model, Qwen 4, is currently in training. That is the entire vendor statement about Qwen 4 as a family. The release's own bullet line says the same thing in shorthand.

Two things follow from that sentence, and both matter more than the sentence does. The first is that "in training" is a status, not a schedule — it tells you work is happening and nothing at all about when it will be finished. The second is that the parameter figure travelling next to Qwen 4 in search results and headlines does not belong to it. Alibaba attaches "5 to 10 trillion parameters" to the Qwen 4.5 and Qwen 5 series, the generations after Qwen 4, in the same release and the same paragraph. No parameter count has ever been published for Qwen 4 itself. That distinction has its own page on this site and is not re-argued here.

The four tier names, and exactly who named them

Reading "Qwen 4" as a single model is the most common error about it. Alibaba names capability tiers within a generation, and four names circulated around the September conference: Max for the flagship, Flash for the high-throughput tier, Plus for the balanced middle tier, and 27B for the size that historically ships as open weights.

The attribution on those four names is worth being precise about, because it is weaker than the vendor's own release and stronger than a rumour. They do not appear in Alibaba's written press release. They were presented from the conference stage and reported out of that session by conference coverage in the days around 22 September 2026. So the correct description of each is: a tier name Alibaba has used in public, with no specification attached to it. Nobody outside Alibaba has a parameter count, a context window, a modality list, a licence or a price for any of the four, and an announcement of a tier name is a statement about a plan rather than a statement about a model.

One spelling deserves a note because readers use it and it can mislead. A large block of demand sits on qwen 4.0 — 1,678 impressions over the period we measured. There is no product called Qwen 4.0 in Alibaba's material and no such model id anywhere. It is a version-string spelling of the same subject, and Qwen 4 Flash is the qualified tier spelling with the most demand behind it. Both route here.

The architecture is the only part of the Qwen 4 question you can actually run today

Alibaba has already published the design it says the Qwen 4 generation will be built on, under a different name — and it is the single most concrete thing anyone outside the company has about the family.

Qwen3.8-Flash-Next landed on the Qwen organisation on Hugging Face on 24 August 2026, and its own model card opens by describing it as "this experimental preview of the architecture that will underpin Qwen4". Its configuration file goes further than the prose: the repository declares model_type as qwen4_exp and its architecture as Qwen4ExpForConditionalGeneration. That string is where the community name "Qwen4-Exp" comes from — it is a configuration value from the vendor, not a product name, and it is not a model you can buy.

What the card says the preview is, in its own figures, all vendor-reported and none of them independently reproduced:

• 125 billion parameters with about 6 billion activated per step, plus a 51-billion-parameter n-gram embedding and a 4-billion-parameter multi-token-prediction head

• QSA, Qwen Sparse Attention, which selects at micro-block level rather than token level and is presented as the mechanism behind lower long-context latency

• Gated residual connections, described as a training-stability measure over widened residual streams

• N-gram embeddings, presented as a cheaper axis for scaling parameters than mixture-of-experts and one that can be offloaded

• A 262,144-token native context, extensible to 1,000,000, with a declared licence of qwen-community-1.0

Read that as evidence about direction and nothing more. Alibaba has not said that any Qwen 4 tier — Max, Flash, Plus or 27B — inherits QSA, gated residual, n-gram embeddings or the training recipe, and it has published no parameter count, layer layout or context length for any of them. The vendor also names Qwen3.8-Flash as the production version built on Flash-Next, which is the version an API caller can reach today; the preview weights are the research artefact.

The release-date question, honestly

Nothing is announced. No date has been given for Qwen 4 or for any tier of it, and no Alibaba statement commits to a quarter, a conference, or a year. Anyone who tells you otherwise is reading a pattern rather than a source, and it is worth naming the patterns, because they are reasonable and they are still not announcements.

The first is generational cadence on the open-weight line. Alibaba's own organisation tells a consistent story: Qwen3.5-27B published 24 February 2026, Qwen3.6-27B on 21 April 2026, Qwen3.8-27B on 5 August 2026 — gaps of roughly two months, then roughly three. Extrapolate that and a next open-weight generation looks due around now. It is an extrapolation from three data points, on a line that already skipped a number: no repository under the Qwen organisation carries the 3.7 generation at all, so the cadence describes what happened, not what is owed.

The second is the conference rhythm. The Qwen 4 status was disclosed at Apsara, Alibaba's annual flagship conference, on 22 September 2026, which invites the reading that the next substantive update arrives at the next one. That is a guess about a calendar, not a plan Alibaba has published.

The third is silicon. The same press release announced the T-Head Zhenwu V900 accelerator, with mass production and commercial release scheduled for the first quarter of 2027, and a supernode cluster design scaling to 500,000 cards. Training and serving a generation at frontier scale is an interconnect problem before it is a compute problem, so a chip schedule is a legitimate thing to reason from. It is also a hardware date that Alibaba projects for its own product, and reasoning from it to "Qwen 4 ships in 2027" is a chain of inference with no vendor sentence under it. The chip timeline and its serving arithmetic are worked through on the roadmap page, and are not re-derived here.

So the honest answer to when is Qwen 4 out is: nobody has said. What exists is a training status from the vendor, a cadence from its own weight releases, and a hardware schedule from its chip unit. Those are three different kinds of evidence and only the first is a statement about Qwen 4.

Where Qwen 4 sits against the line that is shipping

The gap between the two things is the whole of the practical story. The 3.x line is real: Qwen3.5, Qwen3.6 and Qwen3.8 all have open weights on the vendor's own organisation, and Qwen3.7 exists as served API tiers without a published open-weight repository. Qwen3.8 is the current one, and the 27B member of it is the open-weight, locally runnable, quantisable checkpoint that the Qwen 4 27B tier is presumed — presumed, not promised — to succeed. What Alibaba has claimed about Qwen 4 by comparison is a training status and an architecture preview. There is no benchmark, no price, no speed figure and no head-to-head between Qwen 4 and anything, because one side of every such comparison does not exist. If you want the shipping generation's actual numbers, they are on the Qwen 3.8 pages rather than here, and this page will not restate them.

If you need an answer you can act on today

Then the answer is a Qwen 3.x model, and the choice is a real one rather than a consolation. Our own catalogue was read live for this page on 13 October 2026 and returns 207 model identifiers, of which the Qwen card carries the whole shipping line and none of the Qwen 4 line — there is no identifier matching qwen4 or any Qwen 4 tier under any spelling, and this page will not imply one is coming. What the card does carry, under the exact ids as they appear in the list: qwen/qwen3.8-max and its dated snapshot qwen/qwen3.8-max-0902, qwen/qwen3.8-flash, and qwen/qwen3.8-27b, with the earlier generations behind them — qwen/qwen3.7-max, qwen/qwen3.7-plus, qwen/qwen3.7-flash, qwen/qwen3.6-flash, qwen/qwen3.6-plus, and the 3.5 sizes down to qwen/qwen3.5-27b.

That mapping is more useful than it looks, because the two questions readers actually hold are separable. If the question is which tier to use, the answered version is the 3.8 tier list, and Alibaba's own model page puts Qwen3.8-Max at $2 per million input tokens and $6 per million output with a 1M-token context, Qwen3.8-Flash at $0.15 and $0.47, and Qwen3.8-27B at $0.50 and $3.00 — Alibaba's published numbers, not ours. If the question is which generation to plan around, the answer is that planning around Qwen 4 is not yet a decision anyone can make, and running the model the tier will replace is the cheapest way to know whether a Qwen 27B-class dense checkpoint already covers your workload.

There is one structural reason to settle the routing question before the model arrives rather than after, and it is the pattern this family has established. When a 27B open-weight drop lands, the week it lands is the week every endpoint serving it is under load, and the community conversions follow within days. More than 200 models sit on one OpenAI-compatible endpoint at the provider's list price with nothing added, so a Qwen 4 tier would be a model-id change on an existing key rather than a new vendor contract, a new bill and a new SDK. Automatic failover is the part that matters for an unproven checkpoint: a model you are still evaluating belongs behind a fallback chain rather than on a critical path, which is what lets you try the successor the week it appears without betting a production route on it.

What is not settled

Three things are genuinely open and should be named rather than hedged. Whether the Qwen 4 tiers inherit the Flash-Next architecture is unknown — the vendor describes the preview as the architecture "that will underpin Qwen4", which is a statement about intent, and no tier has a configuration file. Whether the 27B tier ships under open weights is unknown, and it is the assumption most readers are actually making, resting on a 3.5-through-3.8 track record rather than on anything Alibaba has said. And whether the four tier names survive to release at all is unknown; the last generation whose tiers were renamed between announcement and launch is not something this page can establish, and neither can anyone else.

Where to read next

The Qwen 4 27B tier is the one most readers are waiting on, and everything published about that specific name — where it sits, and the explicit list of what does and does not exist for it — is on that checklist rather than here. The 5-to-10-trillion-parameter band and the chip timeline that gates any frontier-scale run belong to the Qwen 4.5 and Qwen 5 roadmap page, which is where the serving arithmetic is worked out. And the September conference itself, including what Alibaba showed on stage and what it shipped in the same session, is covered on the Qwen 4 Max lineup page. This page is the map of the three; each of those is the detail.