Qwen3.8-Flash-Next vs Qwen3.8-Max: how close does a 6B-active Qwen4 preview get to the 2.4T flagship?

On August 26, 2026, Alibaba's Qw​en team open-sourced Qwen3.8-Flash-Next, a 176-billion-parameter multimodal mixture-of-experts that activates only about 6 billion parameters per token and is explicitly framed as an early preview of the Qwen4 architecture. The yardstick it will be measured against from day one is Alibaba's own flagship, Qwen3.8-Max, the 2.4-trillion-parameter sparse MoE that reached general availability on August 3 and has been billing $2.00 per million input tokens and $6.00 per million output tokens since. On paper the two sit more than an order of magnitude apart in stored parameters — 2.4 trillion against roughly 180 billion — and roughly sixteenfold apart in active parameters. On Alibaba's own benchmark sheet, the results land far closer than the size — which is either the most impressive thing about the preview, or the clearest reminder yet that a vendor table is a hypothesis until an independent lab runs it.

The short version: this is not a contest between equals, and pretending otherwise is the one way to get it wrong. Qwen3.8-Max is the capability ceiling of the current generation — independently scored by Artificial Analysis, open-weight since August 12, production-proven for three weeks. Qwen3.8-Flash-Next is a one-day-old architecture prototype that Alibaba says most workloads will not need Max's headroom to run. The reason the comparison is still worth making is that on the benchmarks where both models appear in Alibaba's own tables, the preview lands within about a point of the flagship on agentic and office work and ahead of it on two — JobBench and RealWorldQA. The decisive difference is not the quality gap, which is small and unverified. It is the evidence gap: one of these models has a third-party scorecard, and the other has none.

What each model actually is

Qwen3.8-Flash-Next is the architecture trial run. It stores roughly 180 billion parameters on disk — a 125B MoE body, a 51B N-gram embedding table that lives in host memory rather than VRAM, and a multi-token-prediction head — and routes each token through about 6B of them, a sparsity near 95%. Its 512-expert MoE alternates three Gated DeltaNet layers with one Qw​en Sparse Attention layer per group, adds a gated-residual stream and N-gram embeddings, and trained with the Muon optimizer at a cost Alibaba puts at about one-ninth of Qwen3.7-Plus. It takes text, image, and video input, serves a native 262,144-token context extendable to 1M via YaRN, and shipped under the qwen-community-1.0 license on Hugging Face and ModelScope in bf16 and FP8. The production model that will carry the price on the QwenCloud API — Qwen3.8-Flash, announced at $0.16 per million input tokens and $0.47 per million output tokens — is the serving-tuned version of this same architecture and has not opened yet.

Qwen3.8-Max is the generation's finished product. It is a 2.4T sparse MoE on the current-generation architecture — Gated DeltaNet plus gated attention, 512 experts with 10 routed and one shared, 95 billion activated per token per the official model card. It is multimodal on the API, serves a 1M-token context flat at $2/$6 with no long-prompt surcharge, and its open weights, Qwen3.8-2.4T-A95B, landed August 12 under a custom license that is not Apache. It is the model Artificial Analysis has scored at an Intelligence Index of 58 and a Coding Index of 71.8, the strongest independent card any Qw​en 3.8-family model holds.

• Total params — Qwen3.8-Flash-Next: 125B MoE + 51B N-gram (~180B stored). Qwen3.8-Max: 2.4T MoE.

• Active per token — Qwen3.8-Flash-Next: ~6B (under 5%). Qwen3.8-Max: ~95B (~4%).

• Architecture — Qwen3.8-Flash-Next: Qwen4 preview (QSA + GDN, gated residual, N-gram, Muon). Qwen3.8-Max: Qw​en 3.5 generation (GDN + gated attention).

• Context — Qwen3.8-Flash-Next: 262K native / 1M via YaRN. Qwen3.8-Max: 1M on the API / 262K native on the open weights.

• Price — Qwen3.8-Flash-Next: $0.16 / $0.47 per 1M (announced, API not yet open). Qwen3.8-Max: $2.00 / $6.00 per 1M, flat.

• Open weights — Qwen3.8-Flash-Next: qwen-community-1.0, August 26, single-node-capable. Qwen3.8-Max: Qwen3.8-2.4T-A95B, custom license, multi-node.

• Independent score — Qwen3.8-Flash-Next: none yet. Qwen3.8-Max: AA Intelligence Index 58, Coding Index 71.8.

The scoreboard above compresses the comparison to the seven rows that matter, and the sourcing footer carries the load: every Flash-Next figure is Alibaba's own and unreproduced, while the Max column rests on Artificial Analysis plus Alibaba's published table. Read it that way and the columns stop being a horse race and start being a statement about evidence.

The benchmark sheet, read honestly

On the shared rows of Alibaba's own tables, the gap is far smaller than the fortyfold parameter difference would suggest. On CoWorkBench (agentic office work) the preview posts 73.9 against Max's 74.8; on AndroidWorld 84.5 against 85.3; on Agents' Last Exam 51.2 against 52.4 — a point or less on each. On JobBench the preview's 55.7 and on RealWorldQA its 88.5 are the best verified rows on the sheet, ahead of Max. The widening is on the vision-heavy block — LVBench, ERQA, MathVision, Vision2Web — where Max leads by roughly four to six points. On the headline agentic numbers, Qwen3.8-Flash-Next records SWE-bench Pro 62.5 and DeepSWE 58.7, within sight of the flagship on work that costs twelve times more to run.

Now the label, applied to every number above: all of it is Alibaba-reported, from Alibaba's own harness, one day old in the preview's case and unreplicated for both models. That matters differently on each side. Qwen3.8-Max's quality table is vendor-claimed, but its Intelligence Index 58 and Coding Index 71.8 are third-party measurements on Alibaba's API — the model has been exposed to independent scrutiny for weeks, and it holds up. Qwen3.8-Flash-Next has no independent score of any kind yet. The honest reading of the closeness is not "the preview is basically as good as the flagship." It is "Alibaba says the preview is within a point of the flagship, and nobody outside the lab has confirmed it." The gap between those two sentences is the entire difference in how much risk each model carries.

The price gap is the point

Do the arithmetic first, because it is the part that will not move. At $0.16/$0.47, a million input tokens plus a million output tokens bills $0.63 on the production Flash API; the same call on Qwen3.8-Max bills $8.00 — roughly 12.7× more, on both sides of the ledger. Alibaba says the architecture trains at about one-ninth the cost of Qwen3.7-Plus, which is the structural reason the price can sit there, and at $0.16 the production Flash undercuts even DeepSeek-V4-Flash by roughly two-thirds. For high-volume, agentic, or long-context workloads, a 12.7× price gap is the difference between a profitable product and a subsidy — which is why the comparison is being made at all.

This is where a routing layer earns its keep on both sides of the ledger, and the facts are worth stating precisely. Qwen3.8-Max is live on OrcaRouter today at provider list price with zero markup — the same $2.00/$6.00 you would pay the vendor directly, behind the OpenAI-compatible endpoint you already use. Qwen3.8-Flash-Next is not on any API yet: until the production Qwen3.8-Flash opens on QwenCloud, the open weights are the only path, self-hosted. When that API does open, the same 0%-markup pass-through means its $0.16/$0.47 price appears on our side the same day — a price cut that is live in a config, not a line in a press release — and both models sit behind one key with automatic failover between them. You do not rewire the integration to test a day-old architecture; you change a routing rule.

Preview architecture versus proven product

There is a longer-horizon difference that outranks any single benchmark. Qwen3.8-Flash-Next is built on the architecture Alibaba says Qwen4 will use — sparse attention, gated DeltaNet, gated residual, N-gram embeddings, Muon — shipped early as open weights so the ecosystem can stress-test it before the expensive model is built on top. On Alibaba's measurements that stack reaches up to 7.6× faster prefill and 4.9× faster decode at 1M-token context than full attention (vendor-reported), which is the technical reason a 6B-active model can keep a 1M-token window affordable at all. Qwen3.8-Max is the mature product of the current generation, and that is exactly its virtue: the known quantity, the model whose serving behavior and failure modes teams have already mapped.

For a production system the fork cuts both ways. Max is the safe call today — independently scored, three weeks of production exposure, an established price. Flash-Next is the direction Alibaba is betting the next generation on, at a price that reprices the family. Calling it now means riding an architecture the community is still auditing; ignoring it means betting the cheaper architecture does not leapfrog the flagship in six months. Neither bet has to be all-or-nothing — which is the practical point of keeping both reachable from one endpoint.

The two captures above are the public surfaces of each release: the OrcaRouter listing where Qwen3.8-Max is callable today at list price, and Alibaba's Qw​en Studio announcement for the Qwen3.8-Flash-Next open-weights drop.

The decision

Choose Qwen3.8-Max if you need the generation's capability ceiling today — long-horizon agentic work, the strongest independent scorecard in the family, and a model that has been in production for weeks rather than hours. At $2/$6 flat across a million tokens it is cheap for a frontier, and it is the only one of the two with data you did not take on faith.

Choose Qwen3.8-Flash-Next if your workload is high-volume and price-sensitive — agentic loops, batch office automation, long-context passes where a 12.7× price gap is the difference between break-even and margin — and you are comfortable running a one-day-old architecture on a vendor sheet, or self-hosting open weights while the API is pending.

For everyone in between, the sane posture is both: Qwen3.8-Max on one key now, Qwen3.8-Flash-Next a routing rule away when the API opens, and failover so the day-old model cannot take down a production path while you decide. The open question is the honest one — can a model that activates 6 billion of its parameters keep these benchmark numbers under independent replication, or will the vision gap and the one-day-old sheet widen into the real story? The next few weeks answer it, and the comparison is worth revisiting the moment an independent score for the preview lands.