Qwen 4 27B: what Alibaba has confirmed, what it has not, and what to run instead

Qw​en 4 27B is a named member of Aliba​ba's next Qw​en generation that you cannot download, call, price or benchmark today. The Qw​en team showed the name on stage at its Apsara Conference in Hangzhou on 22 September 2026, alongside Qw​en 4 Max, Qw​en 4 Flash and Qw​en 4 Plus, and Aliba​ba's own written release from that day says only that "Qw​en 4 is currently in training". There is no model card for Qw​en 4 27B, no repository, no API identifier, no licence, no context window, no parameter count and no release date. This page is the reference for the name: what the tier is, where it sits in the Qw​en 4 family, and exactly which facts about it exist and which do not.

One correction belongs at the top, because the model has been filed in circulation as "launched" and Aliba​ba's own material does not support that reading. As of 13 October 2026, Qw​en 4 is announced and in training. Nothing in the Qw​en 4 family has shipped, and the only Qw​en generation you can actually run is Qw​en 3.8. That gap — between a name people are already typing and a model that does not exist — is the reason this page exists rather than another announcement post.

Where the name comes from, and how firm it is

The firm part is short and comes from Aliba​ba directly. In its press release "Aliba​ba Unveils Roadmap on Full-Stack AI Strategy from Chips, Cloud Infrastructure, Models to Agents", dated 22 September 2026, the company states that its next-generation model, Qw​en 4, is currently in training, and that the Qw​en 4.5 and Qw​en 5 series ahead of it are projected to scale to 5 to 10 trillion parameters. That is the whole of the vendor's written statement about Qw​en 4 as a family — one sentence, no tiers, no specifications.

The tier names are firmer than a rumour and looser than a datasheet. Qw​en 4 Max, Qw​en 4 Flash, Qw​en 4 Plus and Qw​en 4 27B were presented from the conference stage by the Qw​en LLM lead, and reported out of that session in the days around 22 September. They are not carried in Aliba​ba's written release, and no specification was attached to any of the four on stage. An announcement of a tier name is a statement about a plan. It is not a statement about a model's parameter count, its context, its licence or its price, and this page will not treat it as one.

Practically, that means the name Qw​en 4 27B is confirmed as a tier of the Qw​en 4 family, and everything a purchasing or deployment decision turns on is unconfirmed. The distinction matters more here than for the flagship tiers, because the 27B tier is the one in this family with an established habit of shipping under open weights — Qwen3.5-27B, Qwen3.6-27B and Qwen3.8-27B all exist on the vendor's own Hugging Face organisation — so an announced 27B tier is read by a lot of people as a promise of downloadable weights. It is not one yet. The 27B-named repositories you can actually fetch from that same account today are Qwen3.5-27B, Qwen3.6-27B and Qwen3.8-27B, and the most recent anything under the Qw​en 4 architecture is the Qwen3.8-Flash-Next preview.

The number that gets attached to this family and does not belong to this model

The 5-to-10-trillion-parameter figure travels next to Qw​en 4 27B in search results and gets repeated as though it described the model. It does not. Aliba​ba attaches it to the Qw​en 4.5 and Qw​en 5 series — generations that come after Qw​en 4 — and the same release gives no parameter count for Qw​en 4 itself, let alone for the 27B tier.

This is worth stating on the reference page rather than leaving to a news post, because it is the single most common error in circulation about this family, and because a parameter count attached to a model with no published architecture cannot be checked in either direction. If you see Qw​en 4 27B described as part of a multi-trillion-parameter programme, the number has been borrowed from a different sentence about a different generation.

The architecture is the one part of Qw​en 4 that is genuinely public

Aliba​ba has already published the design it says the Qw​en 4 generation will be built on — not as a Qw​en 4 model, but as a shipped preview under another name. Qwen3.8-Flash-Next landed on the vendor's Hugging Face organisation on 24 August 2026, and its model card states plainly that it is "an experimental preview of the architecture that will underpin Qwen4". That sentence is the most concrete thing anyone outside Aliba​ba has about how this family is built.

The card names four components, all vendor-reported and none of them reproduced by an independent party:

• Hybrid attention with QSA — the Gated DeltaNet and Gated Attention pairing reworked so that the sparse-attention component, Qw​en Sparse Attention, selects at micro-block level rather than token level, which the card presents as the mechanism behind lower long-context latency

• Gated residual — an element-wise data-dependent read gate and a per-branch scalar write gate over widened residual streams, presented as a training-stability measure

• N-gram embedding — a separate parameter store indexed by short n-grams, presented as a cheaper axis for scaling parameters than mixture-of-experts and one that can be offloaded

• A tailored training recipe — Muon and AdamW applied to specific weight categories, with batch-size warmup dropped

The same card gives the preview's own shapes: 125 billion main 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, a 262,144-token native context extensible to 1,000,000, and a declared licence of qwen-community-1.0.

Read that as the family's architecture preview and nothing more. Aliba​ba has not said that the Qw​en 4 27B tier inherits QSA, gated residual, n-gram embeddings or the training recipe, and it has not published a parameter count, a layer layout or a context length for the tier. The preview is evidence about direction; it is not a specification of this model, and treating it as one is how a reference page turns into a rumour.

What the last 27B actually was — the only honest baseline

The useful comparison for Qw​en 4 27B is not a competitor and not the roadmap. It is the previous model to carry the same tier name on the same vendor's account, because that is what a reader will actually be upgrading from.

Qwen3.8-27B was published under the Qw​en organisation on Hugging Face in August 2026. Its model card declares Apache 2.0, describes a causal language model with a vision encoder, and gives these figures, all vendor-reported and all from that card unless noted:

• Parameters — 27 billion, dense; 64 layers; hidden dimension 5,120; a hybrid of Gated DeltaNet and Gated Attention blocks with multi-token prediction trained over multiple steps

• Context — 262,144 tokens natively, extensible to 1,000,000

• Modalities — images and video in, text out, with vision-language understanding described as native rather than bolted on

• Thinking control — thinking mode on by default and switchable per request, with reasoning depth tuned through a reasoning-effort setting and reasoning context retained across turns

• Hosted figures — on Aliba​ba's own Qw​en Cloud page for the model, $0.50 per million input tokens and $3.00 per million output tokens, cache reads at $0.05 and implicit-cache input at $0.10 per million, a maximum output of 131K tokens against a 1M-token context, 5M tokens and 5K requests per minute

Two cautions attach to that list. The first is that the model card says the hosted version of Qwen3.8-27B is "coming soon" while the vendor's Qw​en Cloud page for the same model already carries pricing and a working API sample; where the two disagree, the live product page is the newer material. The second is that none of these numbers tells you anything about Qw​en 4 27B. A predecessor's specification sets the bar a successor is judged against, and that is the whole of its relevance. There is no head-to-head between Qw​en 4 27B and any other model at matched harness and matched effort, because one side of that comparison does not exist.

The specification that does not exist

This is the part most summaries fill in with inference. Here is each dimension, with what the vendor has published:

• Release date — not published. No date has been given for any Qw​en 4 tier.

• Parameter count — not published. "27B" is the tier's name; Aliba​ba has not confirmed a dense 27-billion-parameter model, and has not said whether this tier is dense or sparse.

• Architecture — not published for this tier. The Qw​en 4 architecture preview described above is the family's, not this model's.

• Context window — not published.

• Maximum output — not published.

• Modalities in and out — not published. The previous 27B tier being vision-language does not make this one vision-language, and the Qw​en 4 family containing a multimodal preview does not either.

• Licence — not published. There is no card to declare one, and the tier has no repository to attach a licence file to.

• Weights — not obtainable. There is no Qw​en 4 repository on the vendor's Hugging Face organisation, nothing on its model hub, and no download of any kind.

• Price, free tier, rate limits — none published, and none published for any Qw​en 4 tier. There is no endpoint, so there is no price card, no trial, no credit grant and no rate limit to quote.

• Benchmarks — none, vendor or independent. No Qw​en 4 model appears on a public leaderboard, no harness has been run against one, and no date can be attached to a score because there is no score. Any benchmark number you see attached to Qw​en 4 27B is not a measurement of Qw​en 4 27B.

What you can run today, and where

Nothing in the Qw​en 4 family, on any platform. Aliba​ba has published no Qw​en 4 identifier, and there is nothing to route to.

On our own side, that claim is checked rather than assumed. OrcaRouter's live model list, read on 13 October 2026, returns 207 model identifiers and contains no entry matching any Qw​en 4 tier under any spelling. We do not serve Qw​en 4 27B, and this page will not imply otherwise. What the same list does carry is the previous generation: Qwen3.8-27B, Qwen3.8-Flash, Qwen3.8-Max and the Qwen3.8-Max-0902 snapshot, alongside more than 200 other models on one API.

That is the honest answer to "how do I get a response from Qw​en 4 27B today": you cannot, and the nearest thing that answers the same workload is the model the tier will replace. If the point of the exercise is to find out whether a Qw​en 27B-class dense model fits your work, running Qwen3.8-27B costs a call and settles the question now; the Qw​en 4 tier settles nothing until it has a card.

There is also a forward-looking reason to keep the routing question separate from the model choice. We route more than 200 models through one API, at the provider's list price with nothing added, so an Aliba​ba price change lands on your key the same day rather than at the next billing cycle, and automatic failover means a model you are still evaluating sits behind a fallback chain instead of on a critical path. That structure is worth having in place before a new tier arrives, not after — the week a 27B open-weight drop lands is the week every endpoint serving it is under load.

What would change this page

A reference page for an unshipped model has a short list of events that would make it wrong, and it is more useful to name them than to hedge. Any one of the following would turn Qw​en 4 27B from an announced tier into a model: a model card on the Qw​en organisation, weights in any format, a declared licence, a published parameter count and layer layout, a context and maximum-output figure, a modality list, an API identifier, a price, or a date.

Until one of those exists, the correct description of Qw​en 4 27B is the one Aliba​ba gave: a named tier of a generation that is in training. Everything else on the stem is inference, and the inference is doing a lot of work right now. The Qw​en 4 Max tier is the model most of that inference actually describes, and the Qw​en 4.5 and Qw​en 5 roadmap is where the parameter numbers come from; neither of them is this model.

The one thing worth doing before the weights land is deciding what you would need to see to move. Write down the context length, the licence and the price that would justify a migration off Qwen3.8-27B, and check them against the card when it appears. A published spec sheet is a much easier thing to read than a roadmap, and this tier has a track record of shipping, which is a reason to watch it rather than a reason to plan around it.