Qwen4-Max vs Qwen 4: One Is a Tier, the Other Is a Family, and Only One of Them Is Callable

This comparison is usually written as though it were a fight between two models. It is not. Q​wen 4 is the name of a generation — four tiers A​libaba presented from the stage at its Apsara conference in Hangzhou on 22 September 2026 — and Qwen4-Max is one of those tiers, the flagship, whose preview opened on 11 November 2026 and closes on 25 November. Comparing them is comparing a product to a product line. That is worth doing anyway, because the family name is what people type into a search box and the tier name is what they can actually reach, and the gap between the two is where most of the confusion about this generation lives.

Start with the state of play in one sentence each. Qwen4-Max is real, callable, metered in Credits through an A​libaba subscription, and scheduled to be withdrawn in fourteen days. Q​wen 4 as a whole is one tier in preview and three tiers that exist only as names: Q​wen 4 Flash, Q​wen 4 Plus and Q​wen 4 27B have never had a model card, a price, a context window or a benchmark published for them.

The family, tier by tier, with what exists

A​libaba's own written release from 22 September 2026 says that Q​wen 4 is "currently in training" and attaches no specification to any tier. The four names came from the conference stage, spoken by the Q​wen LLM lead, and were reported out of that session. Tier names announced on a stage are a statement about a plan. Here is the plan against the delivery, as of 11 November 2026:

• Q​wen 4 Max — flagship tier. Delivered as a preview, model ID qwen4-max-preview, 11 to 25 November 2026, Token Plan subscription only, Singapore region. 1M context, 393,216-token output budget shared with thinking, text, image and video in.

• Q​wen 4 Flash — high-throughput tier. Nothing published. No model ID, no card, no date.

• Q​wen 4 Plus — balanced middle tier, historically the broadest multimodal spread in the family. Nothing published.

• Q​wen 4 27B — the open-weight tier, the one this family has a habit of publishing on Hugging Face. Nothing published; no repository exists under the Q​wen organisation.

• The parameter figure — the 5-to-10-trillion-parameter roadmap that travels next to the Q​wen 4 name belongs to the Q​wen 4.5 and Q​wen 5 generations, in the same release and the same paragraph. A​libaba has published no parameter count for Q​wen 4, and none for Qwen4-Max.

One row of four has shipped, and it shipped with an expiry date. That is the whole of the family's delivery record, and it is a better description of "Q​wen 4" than any spec table you will find, because everything else in those tables was invented by someone filling a blank.

What the preview proves about the rest of the family

A shipped flagship is evidence about its siblings even when the siblings are unshipped, and in this case the evidence is unusually specific. Qwen4-Max arrives with three properties that read like family-level decisions rather than tier-level ones.

The first is the 1M context. Max ships with a million-token window; the Q​wen 4 architecture preview — Qwen3.8-Flash-Next, published on Hugging Face on 24 August 2026 with a card describing it as an experimental preview of the architecture that will underpin Q​wen 4 — reports a 262,144-token native context extensible toward 1 million. A family that lands on 1M at the flagship is a family whose other tiers will be described against that number, and the extension story is already written down in the preview's card.

The second is thinking by default at xhigh effort. Max ships with reasoning on, adjustable through reasoning_effort to low, medium or xhigh, defaulting to the top setting. That is a statement about how A​libaba expects the family to be used: the model thinks unless you tell it not to, and the token budget for thinking is drawn from the same 393,216-token pool as the answer. A Flash tier built on the same expectation would be a very different product from a Flash tier built for latency, and the eventual card will tell you which it is.

The third is the API surface. Max is compatible with O​penAI Chat Completions, O​penAI Responses and A​nthropic Messages, and it changes the meaning of max_tokens to cover thinking plus answer together — a change A​libaba documents as starting "with qwen4-max-preview", with older Q​wen models keeping the old behaviour. When a vendor makes the newest model the first one with a new accounting rule, the rule usually travels forward across the generation rather than staying with one tier.

None of this is a specification of Q​wen 4 Flash, Q​wen 4 Plus or Q​wen 4 27B. It is the direction the generation is pointed, from the parts the vendor has actually published, and it is the most anyone outside A​libaba can honestly say about the other three.

The 27B question is the one that matters to most readers

The Q​wen organisation's habit is the reason the 27B tier attracts more attention than a mid-size model usually would. Qwen3.5-27B, Qwen3.6-27B and Qwen3.8-27B all exist as downloadable repositories; Qwen3.8-27B, published in August 2026, declares Apache 2.0, describes a 27-billion-parameter dense model with a vision encoder, a hybrid of Gated DeltaNet and Gated Attention blocks, 262,144 tokens of native context extensible to 1 million, and a thinking mode toggled per request.

That track record is why "Q​wen 4 27B" is read by a great many people as a promise of downloadable weights. It is not one. There is no Q​wen 4 repository on the Q​wen organisation today — the newest things there are the 3.8 family, with Qwen3.8-Flash-Next being the closest any public artefact gets to Q​wen 4. The flagship shipping as a gated subscription preview rather than an open release does not contradict an eventual open 27B, but it does mean the open-weight tier has slipped behind the flagship this cycle in a way it has not before.

For anyone whose architecture cannot route through a Singapore subscription, the practical consequence is blunt: no Q​wen 4 tier is usable by you today, in any form, at any price. The generation you can actually run is 3.8, and the flagship of that generation is a genuinely strong model rather than a placeholder.

Pricing a family that has published one price and it is not per token

Qwen4-Max has no per-million-token rate. What it has is a subscription: Token Plan Personal Edition at $6, $10, $18 or $43 a month for 20,000, 45,000, 80,000 or 200,000 Credits, with usage bundles at $15 per 20,000 Credits, a new-user price that runs to 30 November 2026, and a preview promotion that puts consumption at 20% of standard, dropping to 8% between 22:00 and 08:00 UTC+8. A​libaba states plainly that Credits per request are determined dynamically by model type, token usage, thinking mode and tool calls.

Compare that with the tier Qwen4-Max will replace, which has a real card: Qwen3.8-Max at $2.00 input and $6.00 output per million tokens, cache reads at $0.25, on a 1M-token context. The comparison is not close, and it is not meant to be — one is a subscription whose unit cost you cannot compute, the other is a metered rate you can put in a spreadsheet.

There is a reason to care beyond the preview window. If Q​wen 4 Flash and Q​wen 4 Plus arrive on the same subscription billing, then the whole generation is priced as a product bundle rather than as inference, and the buy-versus-build arithmetic changes for every workload that currently assumes per-token metering. If they arrive metered, the preview was a launch campaign. Which of those happens is knowable within weeks and is the single most decision-relevant thing about this family right now.

What to do while three of the four tiers are names

The expensive mistake here is treating a tier preview as a family launch and planning a migration around tiers that do not exist. The cheap move is to keep the routing decision separate from the model decision, so that whichever tier lands first can be adopted as a configuration change.

That is the case for expressing tier preference as policy rather than as code. OrcaRouter's routing DSL — YAML plus CEL, evaluated per request — lets you pin a workload to a named model and swap that name when a better one appears, without touching the application; a fallback chain means a model you are still evaluating sits behind one you trust rather than on a critical path. Today the catalogue returns 217 model identifiers and none of them is a Q​wen 4 tier — we do not serve Qwen4-Max, and this page will not suggest otherwise. What it does carry is the 3.8 generation: Qwen3.8-Max at A​libaba's $2.00 and $6.00 with 0% markup, the dated Qwen3.8-Max-0902 snapshot, and Qwen3.8-Flash for volume. Putting the workload there now, behind a policy that names the model rather than hard-coding it, is the version of "waiting for Q​wen 4" that ships something.

A dated snapshot is worth calling out separately, because it is the part of this that is easy to lose. Pinning to Qwen3.8-Max-0902 means the model answering your requests cannot change under you while a successor is being previewed, extended or withdrawn. A family in flux is exactly when a snapshot earns its keep.

How this resolves

The comparison collapses into three dates. 25 November 2026 decides whether Qwen4-Max becomes a product or a demonstration; a pay-as-you-go endpoint or an extension past that date would answer the question, and silence would answer it too. The next model card decides whether Q​wen 4 is a family or a flagship — Flash and Plus either get specifications or they stay stage names. And a Hugging Face repository decides the 27B question for everyone who cannot use a subscription API at all.

Until then, the accurate sentence is the one that sounds least like a launch post: one tier of Q​wen 4 exists for fourteen more days, three of them exist on a slide from September, and the Q​wen you can rely on today is the one A​libaba shipped in August. Plan on that and the preview becomes a bonus rather than a dependency.