Qwen4 Max vs Qwen4-Max Preview: One of These Names Has Never Shipped

Type "Qwen4 Max" into a search box and you will get two different things back, and only one of them exists. Qwen4 Max is the flagship tier Alibaba named from the stage at its Apsara conference in Hangzhou on 22 September 2026 — a name with no model card, no published price, no context window, no API identifier and no release date attached to it, fifty days after nothing changed. Qwen4-Max Preview is the container that name eventually resolved to: a real endpoint that opened on 11 November 2026, carries a 1,000,000-token context and a 393,216-token output budget, and is retired at 24:00 on 25 November. Comparing them is not comparing two models. It is comparing a product to a promise, and the useful question is what the promise would have to add before the comparison became real.

The naming problem is the whole story

Alibaba has been unusually consistent about this, which is why the confusion is worth correcting rather than tutting at. The Apsara release of 22 September said the Qwen 4 family was in training and named four tiers — Max, Flash, Plus and 27B — without attaching a single specification to any of them. The written record never claimed otherwise. What circulated afterwards did: coverage and social posts flattened "announced" into "released", and the model-tier names travelled further than the sentence that qualified them.

Then, on 11 November, a model called qwen4-max-preview went live in Model Studio. That is the first and only Qwen 4 tier anyone can call. It is not Qwen4 Max. It is the preview of it, sold under a separate identifier, with its own documentation page, its own access restriction and — the detail that settles the question — its own expiry date.

So the two names describe three different things depending on who is typing: the announced tier, the live preview, and the generally available flagship that neither of them is yet. Getting that straight is most of what anyone needs before spending money.

What actually exists, dimension by dimension

No table will render in this CMS, so here is the contrast one line at a time, with the empty side stated as plainly as the filled one.

• Status — Qwen4-Max Preview: live since 10:00 on 11 November 2026, UTC+8, and retired at 24:00 on 25 November vs Qwen4 Max: named on stage on 22 September 2026, never released.

• Identifier — Qwen4-Max Preview: qwen4-max-preview, with qwen4.0-max-preview as an alias (both work; billing renders the 4.0 form) vs Qwen4 Max: none published.

• Context — Qwen4-Max Preview: 1,000,000 tokens vs Qwen4 Max: not stated.

• Output ceiling — Qwen4-Max Preview: 393,216 tokens, shared by thinking and the final answer vs Qwen4 Max: not stated.

• Input — Qwen4-Max Preview: text, images and video vs Qwen4 Max: not stated.

• Reasoning — Qwen4-Max Preview: on by default, reasoning_effort at low, medium or xhigh, defaulting to xhigh vs Qwen4 Max: not stated.

• Tools and structure — Qwen4-Max Preview: function calling, web search, built-in tools, JSON Object and JSON Schema vs Qwen4 Max: not stated.

• Wire format — Qwen4-Max Preview: OpenAI Chat Completions, OpenAI Responses and Anthropic Messages compatible vs Qwen4 Max: not stated.

• Access — Qwen4-Max Preview: Token Plan only, pay-as-you-go explicitly unsupported, plus a consumer surface in Qwen Chat vs Qwen4 Max: no access path exists.

• Price — Qwen4-Max Preview: no per-token rate published; metered in Credits against a subscription vs Qwen4 Max: no rate card.

• Parameters and weights — Qwen4-Max Preview: neither published vs Qwen4 Max: neither published.

• Benchmark scores — both: none, vendor or independent, as of 11 November 2026.

• End date — Qwen4-Max Preview: 25 November 2026 vs Qwen4 Max: no date, which is the entire problem.

The pattern in that list is that the preview column is complete and the other is not, and that the one place they agree is the place people care about most: neither has a score.

The alias is a clue about what is coming

qwen4.0-max-preview is a small thing that rewards a second look. Alibaba documents that both identifiers work and that billing renders the 4.0 form. That is the behaviour of a vendor who intends a versioned family rather than a one-off codename — the 4.0 namespace is being stood up, and the preview is the first tenant in it.

It also tells you what to watch for. When the flagship lands, the identifier that matters will be a qwen4.0-max-shaped string without the preview suffix, and the thing that will distinguish it will not be a new context window — a million tokens is already claimed here — but the three rows the preview deliberately leaves blank on the commercial side: a per-token rate, a pay-as-you-go endpoint, and a date on which it stops being temporary.

Why a free chat tab is not the flagship arriving

The Qwen Chat launch of Qwen4-Max Preview is the most visible thing that happened to this model, and it is the easiest to misread. A model in a chat product feels permanent, generally available and finished. This one is none of those. The window is fourteen days, Alibaba states that no changes will be made during it, and the model is retired when it closes.

What the chat surface does change is who can evaluate. The API route is Token Plan only with a dedicated base URL at token-plan.ap-southeast-1.maas.aliyuncs.com, so an ordinary Model Studio key will not open it, and the plan's terms restrict usage to "interactive use in coding and agent tools, not automated batch scripts". A browser tab is not a workaround for that licence — it is the surface the licence was describing.

If the flagship arrives, what has to change for you

Assume for a moment that Qwen4 Max ships in December with a published rate and a pay-as-you-go endpoint. Three things in the preview are already telling you what that migration will feel like, and two of them are traps.

The first is the token accounting change. Starting with qwen4-max-preview, max_tokens is equivalent to max_completion_tokens and both limit thinking plus the answer together; every earlier Qwen model applied the cap to the answer alone. If the flagship inherits this, and there is no reason to think it will not, then any request you port from Qwen3.8-Max with a fixed max_tokens will be quietly truncated rather than rejected. The failure is silent, which makes it the kind of thing you find in production rather than in a test suite.

The second is preserve_thinking, on by default in Chat Completions. Prior thinking returns in the assistant messages' reasoning_content field, and folding that into content before resending history is the wrong shape. A multi-turn agent built by someone who has not read that line will degrade in a way that looks like a model quality problem.

The third is effort. reasoning_effort defaults to xhigh, so the default configuration is the slow, expensive one. That is a defensible default for a capability demonstration and a questionable one for a production tier; if the flagship keeps it, budget for the latency you did not ask for.

None of that is exotic. It is the ordinary cost of being early, and the reason to write down what you learn now is that a parameter you set correctly against the preview is a parameter you will not have to rediscover against the flagship.

What to run while the name has no product behind it

For anyone whose actual requirement is "a Qwen flagship I can call today and still be calling in March", the answer predates all of this. Qwen3.8-Max has been generally available since 3 August 2026 with a 1,000,000-token context, text, image and video input, at $2.00 per million input tokens and $6.00 per million output. It is the model the Qwen 4 flagship is intended to replace, which makes it the correct baseline for any evaluation you run this fortnight: whatever the preview does better, measure it against that, not against a blank column.

The routing side of this is where a single endpoint earns its keep rather than being a slogan. Qwen3.8-Max is on OrcaRouter's catalogue at the vendor's list price with 0% markup, alongside 216 other model identifiers, so a price change at Alibaba reaches your bill the same day instead of at the next renewal. Qwen4-Max Preview is not on those routes and will not be — it is reachable through the vendor's own API and its chat product, and pretending otherwise would be the kind of claim this blog exists to avoid. What you can do today is compose them: a routing configuration that tries the preview first and falls through to Qwen3.8-Max means an expired endpoint, a truncated response or a refused video turns into a slower answer rather than a failed one. When the flagship does arrive, that same chain is where you will put it — one line changed, no second contract.

Three dates, and what each one decides

25 November 2026 is the only date anyone has committed to: the preview retires. If Alibaba says nothing before then, the safe reading is that the capability demonstration is over and the flagship is a 2027 question.

The date nobody has published is the one that matters commercially. A Qwen4 Max release would have to arrive with a per-token price to be a purchasing decision, and at that point the interesting comparison stops being preview-versus-flagship — the preview will be gone — and becomes Qwen4 Max versus GPT-6.1 Sol, Gemini 4 Argon, Claude Opus 5.5 and the rest of a market that has been metering tokens for months.

The third date is the one to watch in the other direction: if Qwen 4 27B appears on Hugging Face with open weights, the calculus changes for everyone who cannot put a Singapore subscription in their architecture. That tier has historically been the one the Qwen organisation publishes, and it is the only part of this family where the community might get a model it can actually keep.

Until one of those lands, the honest summary is short. Qwen4-Max Preview is callable, large-windowed, video-capable, unpriced and temporary. Qwen4 Max is a name from a conference stage in September with nothing behind it. And the model to build on this quarter is still the one that shipped in August.