Qwen3.8-Flash-Next vs Qwen3.8-27B: the Qwen4-preview MoE against the open-weights workhorse

The surprising thing about Qwen3.8-Flash-Next is not that Alibaba claims a model activating just 6 billion parameters per token beats most of the open field — it is that serving it needs more memory than the 27-billion-parameter dense model it is being compared against. Qwen3.8-Flash-Next, open-sourced on August 26, 2026 as the Qwen4 architecture preview, keeps a 51-billion-parameter N-gram lookup table in system RAM instead of VRAM; Qw​en3.8-27B, the Apache-2.0 multimodal dense model that shipped in mid-August and is the open-weights workhorse of the current Qw​en 3.8 generation, fits on a single 24GB graphics card at 4-bit. On Alibaba's own benchmark sheet the new MoE beats the 27B on 21 of 22 compared benchmarks. On independent evidence — arena placements, a legal-agent study, third-party throughput measurements — the 27B is the only one of the two with any. That combination is the entire decision.

The matchup in one line: two open-weight, text-plus-vision members of the same generation, same 262K native context, both designed to run outside a datacenter — and split by architecture, by license, by price, and by how much of their story is verified. Qwen3.8-Flash-Next is the sparse MoE that previews everything Qwen4 is expected to inherit. Qw​en3.8-27B is the dense model that compresses what Alibaba learned building the 2.4T flagship into something a single GPU can run. Choosing between them is less about capability — Alibaba says the preview is ahead — and more about whether you trust a one-day-old vendor sheet over a model that has already been picked apart by the community.

The scoreboard

Sourcing first: every Qwen3.8-Flash-Next figure below is Alibaba's own, one day old and unreproduced. The Qw​en3.8-27B's headline benchmark claims are also Alibaba-reported, but the 27B carries independent results — human-preference arena placements, a legal-domain study, and AMD throughput measurements — that the preview cannot match.

• Total params — Qwen3.8-Flash-Next: 125B MoE + 51B N-gram + MTP (~180B stored), 6B active. Qw​en3.8-27B: ~27B dense (28B with vision encoder), all active.

• Architecture — Qwen3.8-Flash-Next: Qwen4 preview — Qw​en Sparse Attention alternating with Gated DeltaNet, gated residual, N-gram embeddings, Muon-trained. Qw​en3.8-27B: 48 GDN linear-attention layers plus 16 full-attention layers (3:1), dense.

• Context — both 262,144 tokens native, extendable to 1M via YaRN.

• Modality — both accept text, image, and video input and return text.

• License — Qwen3.8-Flash-Next: qwen-community-1.0. Qw​en3.8-27B: Apache 2.0.

• Price — Qwen3.8-Flash-Next: $0.16 / $0.47 per 1M (announced; production API not yet open). Qw​en3.8-27B: $0.33 / $2.40 per 1M, live today.

• Run footprint — Qwen3.8-Flash-Next: 51B table in host RAM, ~96GB system memory plus any GPU; FP8 ~186GB, Q4 GGUF ~82GB. Qw​en3.8-27B: BF16 ~55.6GB, 4-bit fits a 24GB card.

• Independent evidence — Qwen3.8-Flash-Next: none yet. Qw​en3.8-27B: #1 open model on Arena.ai Image-to-WebDev, #9 overall on Code Arena WebDev, #1 open-weight on Harvey's Legal Agent benchmark, AMD-measured throughput.

On Alibaba's own numbers, the preview wins nearly everything

Alibaba's published comparison gives Qwen3.8-Flash-Next equal-or-better scores on 21 of 22 language and vision benchmarks against Qw​en3.8-27B, and the wins are not cosmetic. SWE-bench Pro 62.5 against 61.7 is a marginal lead, but DeepSWE 58.7 against 42.2, JobBench 55.7 against 33.4, and CoWorkBench 73.9 against 70.7 are real gaps on exactly the agentic and office workloads the flash tier is aimed at. The preview's headline numbers — LiveCodeBench v6 91.9, GPQA Diamond 91.7, MathVision 95.7 — clear the 27B's corresponding rows on Alibaba's table as well. This is the release's thesis stated as data: most of the reasoning and coding capability of the bigger Qw​en 3.8 line, at a fraction of the active compute.

Keep the label attached to that sentence. These are Alibaba's own evaluations, from Alibaba's own harness, one day old in the preview's case and unreplicated for both. The KGP Talkie architecture teardown that ranks for this matchup reaches the same shape of conclusion, but it is working from the same vendor sheet. A vendor table is a promise; nothing in this paragraph has been independently confirmed.

What the 27B has that Flash-Next doesn't: evidence

This is where the matchup stops being lopsided. Qw​en3.8-27B has spent two weeks in the open, and the community has produced three independent data points. On Arena.ai's Image-to-WebDev leaderboard — blind human votes on which of two anonymized outputs a viewer would rather ship — the 27B ranks #1 among open models and #7 overall with a reported score of 1,574. On the sibling Code Arena WebDev board it sits #9 overall at 1,595, the only model in its size class inside the top ten. Harvey, the legal-AI company, and Engram ran a synthetic-law-firm study that put an adapted 27B at the top of their legal-agent benchmark — with the caveat that the model had studied the test corpus beforehand, which makes it a demonstration of fine-tuning headroom rather than a score for the stock weights. AMD published Day-0 throughput measurements: 24.5 tokens/s on a Ryzen AI Max+ 395 and 51.8 tokens/s on a Radeon AI PRO R9700. None of these is a general-purpose quality score — there is still no Artificial Analysis index for the 27B — but each is a third party that actually ran the model.

Qwen3.8-Flash-Next has none of that. Its entire benchmark sheet is Alibaba's, hours old at the time of writing, and no independent lab has reproduced a single row. That is not an accusation; it is the definition of a day-old release. But it is exactly the asymmetry that should govern the choice: the preview is ahead on the only numbers that exist, and every one of those numbers was written by the vendor that wants you to believe it.

The deployment fork

The practical difference between these two models is where the parameters live, and it determines what hardware you need. Qwen3.8-Flash-Next deliberately offloads its 51B N-gram embedding table to host memory, where it can be asynchronously prefetched — that is why it adds 51B parameters of capacity without adding per-token compute. The consequence is that the model will not fit in a 24GB consumer card the way a local Qw​en has fit before. Community estimates put a Q4 GGUF at roughly 82GB total (about 58GB of main weights plus the 24GB lookup table), the FP8 build around 186GB, BF16 around 360GB; the working recipe is a machine with on the order of 96GB of system RAM plus any GPU that runs the active 6B. The payoff is that per-token compute is cheap — the architecture's entire bet — and long 1M-token context stays affordable because the GDN layers compress history instead of growing KV cache.

Qw​en3.8-27B is the model the other way around: dense, so every token runs all 27B parameters, but small enough that the whole thing fits on hardware you already own. The BF16 weights are roughly 55.6GB; at 4-bit it runs on a 24GB card such as an RTX 3090 or 4090, and AMD's measurements show it doing 24.5 to 51.8 tokens/s on AI Max-class and Radeon PRO hardware. If the constraint is a single workstation, the 27B is the one that actually fits; if the constraint is per-token cost at scale, Flash-Next is the one that wins on paper.

The pricing fork

The API prices tell the same story from the other side. Qw​en3.8-27B is live on OrcaRouter today at $0.33 per million input tokens and $2.40 per million output tokens, provider list price passed through at zero markup — the self-host option made callable, no GPU to buy. Qwen3.8-Flash-Next is not routed anywhere yet: the open weights are the only path until the production Qwen3.8-Flash opens on the QwenCloud API at its announced $0.16/$0.47. When that API opens, the same pass-through puts the $0.16/$0.47 price on our side the same day, on the same key as the 27B, with automatic failover between them — so the way to adopt a day-old architecture is not to bet production on it, but to point a slice of traffic at it with the proven model as the fallback. A routing layer can also front a model you serve yourself, which is the practical path to evaluating Flash-Next's open weights today without a second integration.

Which one to run

Choose Qwen3.8-Flash-Next if you want to ride the Qwen4 direction now, if your workload is high-volume enough that $0.16/$0.47 against $0.33/$2.40 is a real number, or if you have the RAM to self-host it and are comfortable on a vendor sheet. Choose Qw​en3.8-27B if you need Apache-2.0 terms, a single 24GB card, a model you can call today, or any degree of independent evidence — it is the only one of the two that has been run by anyone outside Alibaba.

The two captures above are the public surfaces of each half: the OrcaRouter listing where Qw​en3.8-27B is callable today at $0.33/$2.40, and the Qwen/Qwen3.8-Flash-Next model card on Hugging Face.

And the honest close is a question, because the evidence base is one day old: can a model that activates 6 billion of its parameters keep a 21-of-22 sweep under independent replication, or is the N-gram table that makes it serve lean also the thing that fails on real workloads? The 27B has already survived two weeks of community scrutiny. Flash-Next is where the 27B was two weeks ago — which is the best argument for running them side by side rather than choosing.