A generated hero card titled 'Qwen 4 Max vs Qwen 3.8 Max' with the subtitle 'Build on the model that exists, or wait for the one that does not?' and three pill badges reading 'Qwen3.8-Max GA August 3, 2026', '$2.00 / $6.00 flat 1M' and 'Qwen 4 Max: no date'. A footer line reads 'Alibaba Yunqi Conference, Hangzhou'. Flat vector editorial styling on a white background with a blue-to-cyan gradient wash and the OrcaRouter logo composited bottom right.
Guides & Insights

Qwen 4 Max vs Qwen 3.8 Max: should you build on the model that exists, or wait for the one that does not?

Author

Alistair Wren

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

This is the only matchup in the Qwen 4 Max launch cycle where both sides of the comparison come from the same lab, and it is therefore the only one with a decision attached to it. Qwen 3.8 Max has been generally available since August 3, 2026 at $2.00 per million input tokens and $6.00 per million output tokens with a flat 1,000,000-token window, published weights and a live endpoint. Qwen 4 Max was announced at the Yunqi conference in Hangzhou on September 22, 2026 as the flagship tier of a four-model family that is still in training, with no release date, no price and no published score. The question a reader actually has is not which is better — that cannot be answered — but whether anything about the announcement should change what they build this month. For almost everyone, the answer is no, and the reasoning behind that is more interesting than the conclusion.

What you can buy today, stated precisely

Qwen3.8-Max is a 2.4-trillion-parameter mixture-of-experts with 95 billion parameters active per token. It has been generally available since August 3, 2026, priced at $2.00 input and $6.00 output per million tokens, flat across the full million-token context with no long-prompt surcharge and no threshold at which the rate steps up. Its weights were published on August 12, 2026 as Qwen3.8-2.4T-A95B in BF16 and FP8 under a custom Qwen licence, which means it is inspectable, quantisable and self-hostable by organisations with the hardware to serve it. It records 56 on the Artificial Analysis Intelligence Index at $1.14 per task under the revision in force when it was measured — a revision that has since been superseded twice, so that figure is a historical reading and should be labelled as one whenever it is compared to a current number from a different model.

On our own model page the service profile is concrete: a p50 time-to-first-token of 3.29 seconds, text, image and video input with text output, and 31.3 million tokens of traffic over the trailing seven days against a dated snapshot, qwen3.8-max-0902, carrying the request. That last detail is worth noticing. Alibaba is shipping dated snapshots within the 3.8 generation, which means the generation is still moving underneath a stable model name.

Qwen 4 Max has none of this. What the announcement supplied was a roster and an architecture lineage:

• The four tiers — Qwen 4 Max as flagship, Qwen 4 Flash for throughput, Qwen 4 Plus as the balanced middle, and an open-weight Qwen 4 27B

• The status — described on stage by Qwen LLM lead Liu Da Yiheng as training on a new-generation architecture, with a release described only as soon

• The architecture preview — Qwen3.8-Flash-Next, open-sourced in late August 2026 and labelled on its own card as a Qwen 4 design preview: 125 billion main parameters plus 51 billion N-gram embedding parameters, roughly 6 billion active per step, a 262,000-token native context extending toward 1 million, and a reported training cost roughly 90 percent below Qwen3.7-Plus

• The roadmap number that does not belong here — the 5-trillion-to-10-trillion-parameter target that circulated with the coverage describes the Qwen 4.5 and Qwen 5 generations, not this one. No parameter count has been published for Qwen 4 Max at all

A generated scoreboard titled 'Qwen 4 Max vs Qwen 3.8 Max - the scoreboard'. Left column 'Qwen 4 Max' reads announced not released, not published, not published, not published, not published, Qwen 4 27B promised. Right column 'Qwen 3.8 Max' reads GA since August 3 2026, $2.00 per 1M tokens, $6.00 per 1M tokens, 1,000,000 tokens, 2.4T total 95B active, published August 12 2026. Footer reads 'Qwen3.8-Max figures from its published model card; Qwen 4 Max status per Alibaba's September 22 2026 Yunqi conference.'

What a generational upgrade in this line has actually meant

The most useful evidence about what Qwen 4 Max will change is not the announcement. It is the last upgrade, because Alibaba's recent cadence is unusually well documented and unusually fast.

The 3.8 generation arrived with a preview on July 19, 2026, general availability on August 3, 2026, and weights thirteen days later on August 12. That is a full generation from preview to open weights in under four weeks, and it followed the Qwen 3.7 generation closely enough that the Max tier's useful life before a cheaper tier closed the gap on it has been short. Anyone treating the Qwen flagship as a multi-year architectural commitment has been reading the cadence wrong.

What that implies for Qwen 4 Max is more specific than "wait and see":

• The Flash tier will probably arrive first or alongside the Max tier, and on the Qwen 3.8 record the Flash tier is the one that changes unit economics rather than the one that changes capability ceilings

• The 27B open tier is the most predictable delivery on the roadmap, and the Qwen 3.8 generation demonstrated the downstream ecosystem effect within days of those weights landing

• The Max tier's price, when it comes, will be set against a market where Qwen's own previous flagship sits at $2.00 and $6.00 and where a 428-billion-parameter competitor is at $0.30 and $1.20. The tier name promises capability, not a price position

None of that is a reason to wait. It is a reason to expect the generation to move quickly once it starts moving, and to prefer an architecture that lets you follow it without redeploying.

The case for waiting, stated honestly

There are two situations where waiting is the right call, and they are narrow enough to name.

The first is a workload that is genuinely blocked on something the Qwen 3.8 generation cannot do and the Flash-Next preview suggests Qwen 4 will. Long-context economics is the obvious candidate: Qwen3.8-Flash-Next is built around a sparse attention scheme called QSA that Alibaba credits with a large share of its reported 90 percent training-cost reduction, and if the Qwen 4 tiers inherit it, the cost of running very long contexts could fall materially. If your roadmap is a product that only makes sense at a tenth of current long-context prices, waiting is rational. If it makes sense today, it makes sense today.

The second is an organisation already committed to a self-hosted Qwen deployment. If you are planning hardware around a Qwen checkpoint, the Qwen 4 27B tier and whatever follows it change what you buy, and a hardware decision made three weeks before an open-weight tier is announced is a decision made twice.

Everybody else should be building. A model with no published price, no published context window, no published modality list and no date is not a planning input. It is a roadmap signal, and roadmap signals are for sequencing future work, not for pausing present work.

A screenshot of the OrcaRouter model page for Qwen3.8 Max (qwen/qwen3.8-max), captured September 22 2026, showing the model name, the 1M token context, text image and video input with text output, vision and tool badges, and the on-page navigation.

Making the eventual migration cheap

The reason this decision is less fraught than it looks is that the migration cost between Qwen generations does not have to be high, provided you have not welded a model identifier into your application.

OrcaRouter carries qwen/qwen3.8-max and the rest of the Qwen 3.8 family — Qwen3.8-Flash and Qwen3.8-27B among them — on one OpenAI-compatible endpoint at 0% markup, with the provider's list price passed through untouched. Two consequences follow from that for a team watching this roadmap. First, when Alibaba cuts a price or ships a snapshot, the change lands on the same key the same day rather than at a renewal date, because there is no margin layer to absorb it. Second, and more relevant here, the eventual move from a Qwen 3.8 model to a Qwen 4 tier would be a model identifier in a config rather than a new contract, a new SDK and a new bill — which is what makes "wait and see" a decision you can afford to defer rather than one you have to make now.

Automatic failover covers the other risk. A new flagship tier's first weeks are the worst time to route production traffic through it exclusively, and the ability to express that policy — a slice of traffic to the new model, the rest on the incumbent, with automatic failover between them — in the routing DSL rather than in application code is what turns a generation change into a tuning exercise.

What OrcaRouter does not carry is Qwen 4 Max or any Qwen 4 tier. None of them has been released, and the catalogue reflects that rather than anticipating it.

A screenshot of the OrcaRouter models catalogue, captured September 22 2026, showing the filter sidebar (input modalities, context length, input price, status, series), the model card grid and a code sample posting to the OpenAI-compatible endpoint at api.orcarouter.ai.

The decision, in one page

Build on Qwen3.8-Max now if your workload works on a 1M-token window at $2.00 and $6.00 per million tokens with published open weights behind it. That describes most workloads, and the model is live, priced, benchmarked under a stated revision and serving production traffic. Waiting for a tier with no date buys you nothing you can use this quarter.

Hold a decision if you are buying hardware for a self-hosted Qwen deployment, or if your product's economics depend on long-context costs falling by an order of magnitude. In both cases the Qwen 4 roadmap is genuinely load-bearing, and in both cases the specific tier to watch is not the flagship — it is the 27B open release and the Flash tier's price.

Ignore, in the meantime, any Qwen 4 Max specification table published before Alibaba publishes a model card, and any claim that the 5-to-10-trillion-parameter roadmap figure is a Qwen 4 Max specification. The first is speculation wearing a table's clothes; the second is a number borrowed from a different slide and attached to a model nobody has run.

The generational question resolves the same way it did last time. Alibaba ships fast, prices aggressively, and opens weights on a lag. Whatever Qwen 4 Max turns out to be, the sensible posture is a working system on Qwen3.8-Max and an endpoint that can adopt the successor in a config change — which is a position you can hold through any announcement, including this one.

Compared in this article2

Detected from this article · Benchmarks: Artificial Analysis · updated daily