Qwen3.8-27B vs Qwen 3.8: Same Generation, a GPU Versus a Rack

This week Aliba​ba put Qwen3.8-27B on Cerebras's fast-inference lane — the first time the dense 27B has a genuinely fast hosted option — and Artificial Analysis finally indexed both sides of this matchup. Qwen3.8-27B scores 34 on the AA Intelligence Index. Qwen 3.8, meaning the open core most people mean when they write the name with no suffix, scores 40. Those two numbers reset a comparison that the August launch coverage had to run entirely on the vendor's word.

The model behind "Qwen 3.8" as a standalone name is Qwen3.8-2.4T-A95B: the 2.4-trillion-parameter mixture-of-experts weights Aliba​ba published under a custom license. Qwen3.8-27B is the same generation's dense member — roughly 27 billion parameters, Apache 2.0, sized for a single GPU. They share the family name, the native 262K context length, and a launch window. Everything else about them is a study in opposites: one you can own, the other you can only rent. Here is what each one is actually for now that both have independent numbers.

The same generation, opposite shapes

Qwen3.8-27B is a dense model — 27B parameters (28B counting the vision encoder), 64 layers, a hybrid attention stack of 48 Gated DeltaNet linear-attention layers against 16 full-attention layers. That hybrid is why a 27B can carry 262,144 tokens of native context. Qwen3.8-2.4T-A95B is the flagship architecture: 2.4T total parameters, about 95B active per token, 92 layers with 512 experts per layer, and 69 of those 92 layers on linear attention. Same trick, ninety times the footprint.

• Architecture — Qwen3.8-27B: 27B dense, every token lights up all of it. Qwen3.8-2.4T-A95B: 2.4T total / ~95B active MoE.

• Context — both 262K native; the 27B's 1M extension is hosted-only, the 2.4T reaches ~984K measured.

• Input — Qwen3.8-27B: text, image, and video. Qwen3.8-2.4T-A95B: text-only, thinking locked on.

• License — Qwen3.8-27B: Apache 2.0. Qwen3.8-2.4T-A95B: custom Qwen3.8-Max License.

• Weights — Qwen3.8-27B: 55.6GB BF16, quantizes to a ~16GB Q4 build. Qwen3.8-2.4T-A95B: ~2.4TB BF16, a GPU rack, not a desktop.

• Where you meet them — Qwen3.8-27B: self-hosted or rented cheap. Qwen3.8-2.4T-A95B: rented, because nobody sensible owns the rack.

Independent numbers, finally

Every August vs piece about Qwen3.8-27B carried the same caveat: "no independent scores yet." That is no longer true. Artificial Analysis has both models measured as of this week, and the shape of the data is genuinely interesting.

On the Intelligence Index, Qwen3.8-2.4T-A95B scores 40, ranked 4th of 113 in its class. Qwen3.8-27B scores 34 — which puts it first of 140 models in its open-weights 4–40B size class, a genuinely strong showing for a dense 27B. Both are slow by independent measurement: 39.0 output tokens per second for the 27B and 38.1 for the 2.4T, against a class median around 90. Both are verbose, the 27B generating roughly 200M output tokens across the index runs versus a 68M class median. Neither is a headline frontier model; both are workhorses, and the index says the 2.4T is the stronger workhorse by a clear margin.

Pricing on the independent side tells the same story in dollars. Artificial Analysis prices the 27B at $0.50 per million input and $3.00 per million output on Qwen Cloud's hosted build (90% cache discount), and the 2.4T at $2.00 / $6.00 (88% cache discount). Cost per Intelligence Index task: $0.82 for the 27B against $2.16 for the 2.4T. The smaller model is cheaper to run per unit of intelligence — and both are expensive relative to their speed class because both are reasoning models that burn output tokens.

Everything else — SWE-bench Pro 61.7, DeepSWE 1.1 42.2, LiveCodeBench v6 90.3 for the 27B; the 2.4T's 86.6 Terminal-Bench 2.1 and 67.7 SWE-bench Pro — remains vendor-reported, unreproduced, and labeled as such throughout this piece. The AA indexes above are the independent floor under all of it.

What the Cerebras listing changes

On September 12, 2026, the Qwen account announced that Qwen3.8-27B is now running on Cerebras, its catalogue entry live under the banner "Qwen 3.8 27B is now available on Cerebras PayGo." Cerebras lists it at $0.99 per million input and $1.49 per million output on the WSE-class hardware that made the company's name in speed. That gives the 27B something the 2.4T core has never had: a fast lane.

This matters because slow-and-verbose was the one real strike against the 27B from day one. On the independent harness it does 39 t/s; on our own serving telemetry it has been doing ~154 output tokens per second with a p50 first-token latency of 4.48s — and now a second provider is competing on that same axis. The 2.4T, by contrast, has no fast lane at all: at 38 t/s you are paying frontier prices for mid-pack speed, and the only way around it is a rented GPU cluster running the open weights yourself.

Three lanes for the 27B, one for the 2.4T

As of today the dense 27B is rentable three ways, and the price spread is worth reading before you pick:

• Self-hosted lane — Qwen3.8-27B is live on OrcaRouter at $0.33 / $2.40 per million, run on our own infrastructure. Cheapest input in the market for it, ~154 tok/s output, 262K context.

• Vendor lane — Qwen Cloud's hosted build, $0.50 / $3.00 per million with the 1M-token context and a 90% cache discount.

• Fast lane — Cerebras PayGo, $0.99 / $1.49 per million, positioned on speed rather than price.

The 2.4T open core has no equivalent spread. The flagship API that serves the same architecture — Qwen3.8-Max — is $2.00 / $6.00 per million, and that is the number that sticks whether you call it the Max or the open core. Run the weights instead and the economics flip to hardware: the 2.4T needs ~2.4TB of BF16 weights and a multi-GPU node, a capital decision nobody makes casually. A 24GB GPU runs the 27B's Q4 build at a marginal cost that rounds to zero.

The worked example that decides most teams: 100M input and 20M output tokens a day. On the 2.4T's $2/$6 rate that is about $320 a day, ~$117,000 a year. On the 27B at our $0.33/$2.40 it is about $81 a day — and on the self-hosted copy it is the price of electricity. The 2.4T buys you a real quality edge, AA 40 against 34. Whether that edge is worth a five-figure monthly bill is the entire question this pairing asks.

Where they genuinely diverge

Beyond the index, three practical differences decide which one fits a given workload.

Multimodal input is the cleanest filter. Qwen3.8-27B reads text, images, and video — screenshots, charts, documents with figures, video frames all go into the same 262K window. Qwen3.8-2.4T-A95B is aggressively text-only. If your input includes anything visual, the 2.4T is disqualified regardless of its higher index.

Thinking control is second. The 27B's reasoning mode can be switched off per request, which matters for latency-sensitive extraction where a chain of thought is pure overhead; it also exposes reasoning_effort. The 2.4T core cannot turn thinking off — every response opens with a think> block, and you pay for those tokens whether you wanted them or not. The flagship API fixes this, but the open core does not.

Licensing is third, and it quietly matters at scale. Apache 2.0 on the 27B is the permissive standard: modify, redistribute, commercialize, patent grant included. The 2.4T ships under a custom Qwen3.8-Max license — permissive in spirit, but it is Aliba​ba's terms, not a community standard, and it is worth reading the file before you build a product on it.

Routing the decision

Both sides of this matchup are reachable through one OrcaRouter key, which is what makes the "which do I pick" question cheap to answer empirically. Qwen3.8-27B is live as qwen/qwen3.8-27b at provider list price with zero markup, and the flagship API built on the same 2.4T core is live as qwen/qwen3.8-max at $2.00 / $6.00 per million — Aliba​ba's own rate, passed straight through, so any vendor price change is live here the same day.

That same 2.4T architecture as downloadable weights is not something we serve — if you want the open core over an API today, the flagship lane is the practical way to reach it, and qwen/qwen3.8 is pre-wired to switch on the moment a provider serves the open weights. A routing rule that sends the visual and latency-sensitive traffic to qwen/qwen3.8-27b and the hard-reasoning tail to qwen/qwen3.8-max costs nothing to set up and fails over automatically between providers. One key, no second contract, and the split is a configuration change rather than a re-architecture — the cheapest way to test whether the 2.4T's extra six index points are worth $5.67 per million output tokens to you.

Who should pick which

• Pick Qwen3.8-27B if your input includes images or video, if you want thinking you can turn off, if you own a 24GB GPU or plan to, or if your per-token spend is volume enough that $0.33/$2.40 versus $2.00/$6.00 is the whole argument.

• Pick Qwen 3.8 (the 2.4T core) if you are text-only, you want the strongest independent reasoning number in the family (AA 40), and the $2/$6 rate — or the cost of running a GPU node — is comfortably inside your budget.

• Pick both if your traffic spans the two profiles: route through the gateway, let the router split by modality and difficulty, and change your mind when either model's next checkpoint or price lands.

The verdict

The Qwen 3.8 name was always two products pretending to be one family, and now both have independent numbers to argue about. Qwen3.8-2.4T-A95B is the stronger brain, text-only, expensive, and undeniably rentable at $2/$6. Qwen3.8-27B is the deployable one — multimodal, Apache 2.0, self-hostable, and as of this week it finally has a fast lane too. The index gap is real: 40 against 34. So is the price gap: roughly five times the per-token cost when you run the 2.4T as an API. For most teams the decision is not about the six index points. It is about whether you are buying tokens or buying hardware — and the 27B is the only one of the two that lets you buy the hardware.