Qwen4-Max Preview: a 1M-Token Flagship You Can Only Rent for Fifteen Days

Most model launches arrive with a start date. The Qwen4-Max preview arrived with an end date. A​libaba opened it at 10:00 on 11 November 2026 (UTC+8) and said, on the same documentation page that announced it, that the model will be retired after 24:00 on 25 November — fifteen days later, with no changes permitted in between. Inside that window you get a 1,000,000-token context, a 393,216-token output ceiling that thinking and the final answer have to share, and native image and video input. What you do not get is a per-token price, a parameter count, published weights, or a single benchmark number, vendor-reported or otherwise. The launch post says the model is "built for real-world tasks, with deeper knowledge and enhanced capabilities", which is the kind of sentence every launch post contains; the documentation says the model has an expiry date, which almost none of them do.

What actually went live at 10:00

The product name is Qwen4-Max-Preview and the model ID is qwen4-max-preview. There is also an alias, qwen4.0-max-preview, and A​libaba notes that both IDs work but bills render the 4.0 form — worth knowing before you go looking for spend under one string and find it under the other.

There is exactly one route to it, and it is not the ordinary API. Qwen4-Max-Preview is served through Token Plan only; pay-as-you-go access is explicitly not supported. Token Plan Personal Edition is a subscription in Credits, available only in A​libaba's Singapore region, with a dedicated base URL at token-plan.ap-southeast-1.maas.aliyuncs.com. If you have a Model Studio API key that works against qwen3.8-max, it is not the key that opens this model.

The timing is worth a line because the gap is short. A​libaba named the Q​wen 4 family — Max, Flash, Plus and 27B — from the stage at its Apsara conference in Hangzhou on 22 September 2026, and its written release that day said only that Q​wen 4 was in training. Fifty days later the flagship tier has a working endpoint. That is fast for a frontier family and it is the strongest argument for taking the preview seriously rather than filing it as a stage demo.

The specification, as published

This is the whole of it, from A​libaba's own model page rather than from coverage of it:

• Context — 1,000,000 tokens.

• Maximum output — 393,216 tokens, shared by thinking and the final answer.

• Input — text, images and video. Output is text.

• Thinking — on by default, with reasoning_effort set to low, medium or xhigh, defaulting to xhigh.

• Tools — function calling, web search and built-in tools.

• Structured output — JSON Object and JSON Schema.

• API shape — O​penAI Chat Completions, O​penAI Responses and A​nthropic Messages compatible.

• Price per million tokens — not published. The model is metered in Credits against a subscription, and A​libaba does not publish a credits-per-token rate.

• Parameters — not published. Weights — not published. Licence — not applicable, because there is nothing to download.

• Benchmark scores — none. Not a vendor table, not an independent run.

• Availability window — 10:00 on 11 November to 24:00 on 25 November 2026, UTC+8, then retired.

Read the output ceiling twice. 393,216 is a large number and it is not the context window — it is the total budget for a single response, and reasoning tokens come out of it before the answer starts. A model that defaults to xhigh effort will spend a meaningful share of that budget thinking, which makes the effective answer length considerably shorter than the figure suggests.

The line that will break a port: max_tokens

Buried in the usage notes is a change that matters more than most of the spec sheet. Starting with qwen4-max-preview, max_tokens is equivalent to max_completion_tokens, and both limit the combined tokens for thinking and the final answer. Every earlier Q​wen model keeps its previous behaviour, where the cap applied to the answer. So a request ported from qwen3.8-max with max_tokens: 4096 now buys 4,096 tokens of thinking plus whatever is left, which on a hard prompt can be nothing at all — and the failure mode is a truncated response rather than an error you can catch.

Two other defaults are worth setting deliberately. Chat Completions enables preserve_thinking by default, so historical thinking comes back in the assistant messages' reasoning_content field; concatenating that into content and sending it back as history is the wrong shape and will confuse the next turn. And reasoning_effort and thinking_budget must not be set together — pick one control.

None of this is exotic. It is the ordinary cost of being early: the compatibility layer that makes the model look like every other O​penAI-shaped endpoint does not make its token accounting identical, and the difference is only visible when something gets cut off.

What fifteen days of Token Plan costs

Because there is no per-token price, the only pricing question a reader can actually answer is what the subscription costs. A​libaba's four Personal Edition tiers run at $6 a month for 20,000 Credits, $10 for 45,000, $18 for 80,000 and $43 for 200,000, with extra bundles at $15 per 20,000 Credits and a new-user discount that runs to 30 November. Credits are consumed at a rate the vendor says is determined dynamically by model type, token usage, thinking mode and tool calls, and the only place to see the real rate is the usage detail in the console.

The preview carries its own promotion: during the window, consumption runs "as low as 20% of the standard rate", and calls between 22:00 and 08:00 UTC+8 get a further discount that lands at 8% of standard. The two pages that describe that night rate disagree about how to phrase it — the model page calls the extra step 40% off, the Token Plan page calls it 60% off — and both arrive at the same 8%, so treat the arithmetic as the fact and the percentage as copy.

One restriction deserves to be quoted rather than paraphrased, because it rules out most production uses before price is even a question: Token Plan Personal Edition is for "interactive use in coding and agent tools, not automated batch scripts". A preview you cannot batch is a preview you evaluate by hand, in an editor, the way you would evaluate a model in a chat window.

What the missing numbers cost you

Three absences do real work here, and each blocks a different decision.

No price blocks the budget. Credits cannot be converted into dollars per million tokens from anything A​libaba publishes, so a team cannot compare the preview against a metered model on the only axis that matters for volume work. The subscription sets a ceiling on spend, not a rate.

No parameter count blocks the infrastructure guess. The nearest thing to an architecture statement is Qwen3.8-Flash-Next, published under the Q​wen organisation on Hugging Face on 24 August 2026, whose card describes it as an experimental preview of the architecture that will underpin Q​wen 4. That card reports 125 billion main parameters with roughly 6 billion active per step, a 51-billion-parameter n-gram embedding store, and three named techniques — QSA sparse attention, gated residual connections, and n-gram embeddings as a separate lookup. It is vendor-reported and unreproduced, and it describes a 180-billion-parameter preview model, not this flagship. Anyone quoting a parameter count for Qwen4-Max is quoting something A​libaba has not said.

No score blocks the ranking. The Artificial Analysis leaderboard, read on 11 November 2026, carries no Qwen4 entry at all. That is not a criticism of the model — a preview that opened this morning has had no time to be measured — but it does mean every "Qwen4 beats X" figure already circulating is a projection or a different model's number wearing this one's name.

There is also a fourth absence that is easy to miss: "no changes will be made during this period". The preview is frozen. A behaviour you find annoying on day two will still be there on day fourteen, and reporting it buys you nothing until the next model exists. That is the correct trade for a preview and the wrong trade for a dependency.

Sending a request this week, and what to run instead

If you want to see the model, the path is a Token Plan subscription, the dedicated base URL, and the model ID above; the endpoint speaks O​penAI and A​nthropic wire formats, so an existing client needs a base URL and a key rather than a rewrite.

If you want to ship something, this is not the model to ship on, and it is worth being plain about the alternative. OrcaRouter's catalogue, read today, lists 217 model identifiers and none of them is a Q​wen 4 tier — we do not serve Qwen4-Max, and no amount of routing cleverness changes that. What the same catalogue does carry is the model this one will eventually replace: Qwen3.8-Max, generally available, 1M context, text, image and video in, at $2.00 per million input tokens and $6.00 per million output, at A​libaba's list price with nothing added. That last part is the structural point rather than a slogan: because we pass the provider's price through at 0% markup, a vendor price change lands on your key the day it happens — which is exactly the mechanism that will matter on the day a Q​wen 4 tier gets a pay-as-you-go rate, if it gets one.

For the evaluation itself, a fallback chain is the cheap way to run it. Put the preview behind a model you already trust rather than in front of it: if a request errors, times out or truncates because the thinking budget ate the answer, the fallback answers and you still learn whether 1M context and video input solve your problem. Trying an unproven endpoint without a fallback is how a fifteen-minute experiment becomes an outage.

What 25 November decides

Three outcomes are possible and they are not equally likely. The clean one is general availability with a published per-token price and a pay-as-you-go endpoint; that would make Qwen4-Max a real purchasing decision and would make this preview the cheapest possible due diligence. The second is an extension — another window, same terms — which would tell you A​libaba is still capacity-constrained in Singapore and still not ready to meter the model. The third is retirement on schedule with silence, which would mean the preview was a capability demonstration aimed at enterprises buying through Token Plan rather than at developers buying tokens.

Watch two things alongside it. Whether the other three tiers get model cards, because a family with one shipped tier and three names is a roadmap, not a lineup. And whether Q​wen 4 27B appears with open weights, because that tier has historically been the one the Q​wen organisation publishes on Hugging Face, and a downloadable Q​wen 4 model would change the calculus for everyone who cannot put a Singapore subscription in their architecture.

For now the honest summary is narrow and useful: Qwen4-Max exists, it is callable, it has a million-token window and a 393,216-token output budget, it will be gone in fifteen days, and nobody outside A​libaba has measured it. That is enough to spend an afternoon on and not enough to plan a quarter around.