Qwen 3.8 MoE: Which Members Are Mixture-of-Experts, and What It Changes for You

Yes — the 3.8 generation contains mixture-of-experts models, and the ones that answer to the bare question are Qwen3.8-2.4T-A95B, the 2.4-trillion-parameter Max core A​libaba published as open weights in August 2026, and Qwen3.8-Flash-Next, the 125-billion-parameter sparse preview of the Qwen4 architecture. The two members a reader can actually call — Qwen3.8-Max and Qwen3.8-Flash — are the vendor's own served builds derived from those two architectures, so they are sparse by descent even though their internal configurations are not published. The one exception is the member most people arrive here already holding: Qwen3.8-27B is dense, and nothing in its name says so. That is the whole answer, and the rest of this page is the figures behind it, each one taken from A​libaba's own model card rather than from a discussion thread.

The dates in this article are used to date the models, not to announce them. What this page answers is the question readers are typing now: on 2026-09-15 to 2026-10-12 our own Search Console property logged 432 impressions and 35 clicks on the single query qwen 3.8 moe — an 8.1% click-through, the highest on any query in its batch — plus three long-tail rows that take the family to 440 impressions and 36 clicks. A reader typing that is not looking for news. They are trying to find out whether the model they are about to download or call is sparse, because the answer decides what they can run. Nothing on this site answered it, which is why this page exists. Where a question is about a specific member rather than the architecture, the page that owns it is named as we go.

The name tells you, if you know the notation

A​libaba does not put "MoE" in any Qwen 3.8 model name. It writes the parameter count as total-Aactive, so 2.4T-A95B reads as 2.4 trillion parameters stored and 95 billion activated per token. The -A half is the tell: a name with an activation figure in it is a sparse model, and a name with a single round number of billions is a dense one. Qwen3.8-27B carries no activation figure because every one of its parameters is live on every token, and its card calls it "a compact, deployment-friendly dense model" in the first sentence that describes it.

That reading holds across the generation:

• Qwen3.8-2.4T-A95B — mixture-of-experts, 512 experts, 2.4T stored / 95B active, open weights, custom licence.

• Qwen3.8-Flash-Next — mixture-of-experts, 512 experts, 125B stored plus a 51B n-gram table / 6B active, open weights, community licence.

• Qwen3.8-Max — the served build; A​libaba's own card for the weights says Max is "the official version based on Qwen3.8-2.4T-A95B", and its cloud model page describes it as a "2.4-trillion-parameter MoE flagship". Sparse.

• Qwen3.8-Flash — the served build derived from Flash-Next, carrying a 1M-token context by default. Sparse by descent.

• Qwen3.8-27B — dense, 27B stored and 27B active, Apache 2.0. No experts anywhere in its configuration.

The architecture, from the cards themselves

Both sparse members are documented in the same shape on Hugging Face, and the numbers below are quoted from Qwen/Qwen3.8-2.4T-A95B and Qwen/Qwen3.8-Flash-Next respectively. Where a figure is not on a card, this page says so rather than borrowing one from a forum.

Qwen3.8-2.4T-A95B — the Max core

It is a 92-layer causal language model whose block structure is written on the card as 23 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE)). In plain terms: three linear-attention blocks between every full-attention block, and a mixture-of-experts feed-forward layer after both. The expert layer is specified as 512 experts, 10 routed plus 1 shared activated per token, with an expert intermediate dimension of 2048. Stored parameters are 2.4T and activated parameters are 95B. The attention scheme is hybrid: Gated DeltaNet with 128 linear heads for V and 16 for QK at head dimension 128, paired with Gated Attention at 64 query heads and 4 KV heads at head dimension 256, with a rotary dimension of 64. Context is 262,144 tokens natively, extensible to 1,010,000. The card also notes multi-token prediction trained with multiple steps, and the configuration file confirms the sparse fields directly: num_experts: 512, num_experts_per_tok: 10, num_hidden_layers: 92.

Two things the card is explicit about and a reader should not gloss over. It is text-only — no vision tower — and the card's own tip states that the hosted Max adds vision input, non-thinking support, 1M context by default and built-in tools on top of these weights, so the downloadable artifact is narrower than the API it shares a core with. And it ships under a custom agreement, qwen3.8-max, not Apache 2.0 and not the community licence; this page does not re-argue those terms, and they are covered on the page for the weights themselves.

Qwen3.8-Flash-Next — the smaller sparse architecture

This is the member where the sparse ratio is most extreme and most easily misread. Its card gives 48 layers arranged as 12 × (3 × (Gated DeltaNet → MoE) → 1 × (Qwen Sparse Attention → MoE)), the same 512 experts with 10 routed plus 1 shared, and an expert intermediate dimension of 640 — a third of the Max core's. Parameters are stated unusually: "125B with 6B activated, plus 51B n-gram embedding and 4B MTP". Those three components are why its repository lists 180B in the sidebar while the MoE body is 125B, and why a reader comparing "125B vs 2.4T" is comparing the wrong numbers. Attention is Gated DeltaNet paired with Qwen Sparse Attention, which the card says operates at the micro-block level rather than selecting individual tokens. Context is 262,144 natively, extensible to 1,000,000. Licence: Qwen Community License 1.0. Its configuration file also records num_experts: 512, num_experts_per_tok: 10 and 48 layers.

What A​libaba has not published

For the served Qwen3.8-Max there is no expert count, no layer count, no activated-parameter figure and no per-token routing specification in any vendor material this page could find. The vendor describes it as a 2.4-trillion-parameter MoE flagship and states that it is built on Qwen3.8-2.4T-A95B; that is the limit of the published detail, and it is a reasonable inference from those two statements that the architecture matches, not a documented fact. The same gap applies in kind to the served Qwen3.8-Flash, whose derivation from Flash-Next is stated but whose internal configuration is not published at all. Any page that hands you an expert count for the hosted Max is filling a gap the vendor left open.

What the active-parameter figure changes about running it

This is why the question gets asked at all, and the answer runs in two directions at once.

Download size follows the total, not the active count. Qwen3.8-2.4T-A95B is roughly 4.9 TB across 213 shards in its BF16 repository and roughly 2.5 TB in the FP8 sibling — either way a rack-scale artifact, not a workstation one. Qwen3.8-Flash-Next is about 360 GB in BF16 across 131 shards, and about 180 GB by its own sidebar count once the components are added up. On disk, nothing about the sparse design makes either one small; the expert weights all exist and all have to be stored.

Compute and memory bandwidth follow the active count. Because only 95B of the Max core's 2.4T parameters, and only 6B of Flash-Next's stored parameters, participate in any one token, the per-token cost of these models is closer to that of a much smaller dense model than their file sizes suggest. That gap is the entire reason a 2.4-trillion-parameter model can be served at all, and it is why the same total-to-active ratio that looks like a footnote on the card is the most operationally relevant number on it. This page will not put a tokens-per-second figure next to it, because no rate we could state would hold across two different quantisations and an unknown serving stack — the ratio is the transferable fact, the throughput is not.

Flash-Next adds a second wrinkle that belongs here rather than in a footnote. Its 51B n-gram embedding table is a lookup structure, and the card's own framing is that embeddings are "more amenable to offloading than Mixture-of-Experts". That is what makes a checkpoint this size runnable with the embedding table resident in host memory rather than pinned in accelerator memory — a design choice, not a claim about your machine. The practical consequence is that Flash-Next is easier to fit by adding system RAM than by adding VRAM, which is the opposite of how a dense model of equivalent stored size behaves.

Where the dense sibling sits

Qwen3.8-27B is the counter-example the rest of the generation is measured against: 64 layers, a 17,408-wide feed-forward intermediate dimension instead of an expert bank, 27B stored and 27B active, Apache 2.0, and about 55.6 GB of BF16 weights that fit on hardware a person owns. Where the sparse members trade storage for a small per-token compute footprint, the dense one trades per-token cost for predictability — every parameter is live, so memory and speed are trivial to size. The full treatment of that model, its formats and its fit belongs to the Qwen3.8-27B page, which owns it; this page only needs the contrast.

What is not settled

Whether the served Qwen3.8-Max and Qwen3.8-Flash actually instantiate the same expert counts and routing as the downloadable cores is unverified: the vendor states the derivation and stops there, so treat any specific expert count for the API models as unconfirmed. And the open quantisation ecosystem around Flash-Next has produced builds from roughly 1-bit to full precision, with no controlled low-bit comparison published on matched hardware — if you are choosing a bit-width, that call is yours to measure.

Availability

The two members whose expert counts are documented — Qwen3.8-2.4T-A95B and Qwen3.8-Flash-Next — are download-and-host artifacts. Neither is in our model list today, and there is no id resembling them, so this page makes no availability claim about them or about any model named "Qwen3.8-MoE": the name describes an architecture, and nothing in our catalogue is sold under it. What we do serve is three other members of the same generation, and a reader whose question was architectural can call any of them:

• qwen/qwen3.8-27b — the dense, Apache-2.0, self-hostable one; text, image and video in, 262,144-token context, $0.33 per million input tokens and $2.40 per million output on the provider's published list.

• qwen/qwen3.8-flash — the sparse production lane derived from Flash-Next; 1M-token context, $0.15 per million input and $0.47 per million output, with cached input around $0.018.

• qwen/qwen3.8-max — the Max core behind an endpoint; 1M-token context, $2.00 per million input and $6.00 per million output.

All three sit behind one OpenAI-compatible endpoint, so running the dense shape against the served sparse one on your own traffic costs a model-id change rather than an integration. Those prices are the provider's published list passed through with no markup, which is why a vendor price change reaches your endpoint the same day.

The dense half of that comparison is also the escape hatch on hardware, and it is the one most of this page's readers end up taking. Qwen3.8-27B is around 55.6 GB of BF16 weights, 27B stored and 27B active, Apache 2.0 — so it is the only member of this generation whose fit you can size on one calculator. The sparse members are cheaper per token served and heavier to own: roughly 4.9 TB of weights for the Max core, roughly 360 GB for Flash-Next, both needing infrastructure you provision rather than rent. The Qwen3.8-27B page owns the fit, the formats and the licence questions; this page stops at the architecture.

So does the architecture answer change what you should run? Less than the search that brought you here implies. If you are calling a model, the sparse members' saving is already priced into what we charge and the choice is a cost-and-quality one, not an architectural one. If you are downloading one, the answer is that no Qwen 3.8 model with 512 experts fits a workstation — the one that does is the dense 27B, and that is the decision the -A notation was quietly making for you all along.

The dense half of that comparison is also the escape hatch on hardware, and it is the one most of this page's readers end up taking. Qwen3.8-27B is around 55.6 GB of BF16 weights, 27B stored and 27B active, Apache 2.0 — so it is the only member of this generation whose fit you can size on one calculator.