Alibaba opened a preview of Qwen4-Max on November 11, 2026 — the first model of its next-generation Qwen4 family. It comes with a 1-million-token context window, image and video understanding, and deep reasoning. And it comes with a catch most launch coverage buries: the preview ends on November 25, after which Alibaba says the model will be retired. You have fourteen days to try it, and there is still no benchmark, no price per token, and no date for the full release.
That is the honest shape of this launch. Qwen4-Max Preview is real, you can use it today, and it is also a two-week evaluation window for a model Alibaba has barely documented. If you are a freelancer, a developer, or a small business owner wondering whether this is the next model to build on, the short answer is: try it now if you are curious, but do not build anything on it that needs to still work on November 26.
What actually launched on November 11
At 10:00 Beijing time on November 11, the official Qwen account on X posted “Qwen4 is coming.” — and Alibaba listed Qwen4-Max-Preview (model ID qwen4-max-preview, alias qwen4.0-max-preview) on its Token Plan subscription. Early access opened the same day in QwenWork, Qoder, the Qwen app, and QwenStudio, and the model has since appeared in Qwen Chat as well.
This matters because Qwen4 has been the most-watched open-weight-adjacent launch of the fall. Alibaba named the Qwen4 family at its Apsara Conference in Hangzhou on September 22, 2026, when Qwen project lead Liu Dayiheng said the new family was in training on a new-generation architecture. Four tiers were named on stage: Max, Flash, Plus, and an open-weight 27B model. For seven weeks after that, there was no release — only a teaser poster on November 6 showing birds over a dark sea and the line “The becoming begins.”
November 11 changed that, but only partly. Qwen4-Max is the first Qwen4 model of any kind to reach users. The rest of the family — Flash, Plus, and the 27B — still has no release date, no weights, and no model card. Alibaba has not said when the full Qwen4 release arrives, or whether the preview model becomes a permanent product in some form.
One thing to get straight early: this is not an open-weights release. There is nothing to download from Hugging Face, no model card to read, and no weights to run on your own hardware. Qwen4-Max Preview is a hosted model you reach through Alibaba’s own surfaces. The open-weight question — the one the community cares most about, given the 27B tier on the roadmap — is still unanswered.
The two-week window, and what “retired” means
Alibaba’s own Model Studio documentation states the preview window plainly: from 10:00 on November 11 to 24:00 on November 25, 2026 (UTC+8). During that window, Alibaba says no changes will be made to the model. After it ends, the model will be retired.
Retired is not the same as “updated” or “replaced.” It means this exact model version goes away. Anything you benchmark, prompt, or tune against qwen4-max-preview during these two weeks describes a model that will not exist on November 26. That is a strange way to launch a flagship, and it is the single most important fact about this release for anyone deciding how much time to invest in it.
Why would Alibaba do it this way? The company has not said. The most plausible reading is that this is a controlled evaluation: a fixed, unchanging snapshot that developers can test against while Alibaba finishes the full Qwen4 release. A frozen model makes comparisons fair — your results on November 12 and your results on November 24 describe the same thing. But it also means the preview is a demo with an expiry date, not a product launch.
If you are keeping score, the window is genuinely short. From today, November 12, you have thirteen days left. A freelancer who wants to run a serious comparison — say, Qwen4-Max against Claude or GPT-6 on a real client codebase — needs to start now, because the model will not wait for a free weekend.
What Alibaba says Qwen4-Max can do
Here is what Alibaba’s Model Studio documentation actually lists for Qwen4-Max-Preview:
| Spec | What Alibaba’s docs say |
|---|---|
| Context window | 1,000,000 tokens |
| Max output | Up to 393,216 tokens (thinking and the answer share that budget) |
| Inputs | Text, image, and video |
| Reasoning | Thinking on by default |
| Model ID | qwen4-max-preview (alias qwen4.0-max-preview) |
| API compatibility | OpenAI Chat Completions, Responses, and Anthropic Messages APIs |
| Availability | Token Plan only — pay-as-you-go is not supported |
A 1-million-token context window is the headline number, and it is worth translating into plain English. A token is roughly three-quarters of a word, so a million tokens is roughly 700,000–750,000 words of text — something like ten long novels, or a medium-sized codebase, held in the model’s working memory at once. That is the same context size as Qwen3.8-Max, the current flagship it is meant to replace, and it puts Qwen4-Max in the same league as the biggest-context frontier models on raw window size.
But a big window is not the same as a good model, and the output ceiling deserves a second look. The 393,216-token output budget has to cover both the model’s internal thinking and its final answer. On a hard reasoning task, the thinking alone can eat a large share of that budget before a single word of the answer appears. That is normal for reasoning models, but it is a practical limit worth knowing before you paste a huge codebase into it and ask for a rewrite.
Alibaba says the model beats Qwen3.8 on world knowledge, code generation, and sustained performance on long tasks with repeated tool calls. That claim has no published benchmark behind it yet. There is no score on any public leaderboard — the Artificial Analysis leaderboard carried no Qwen4 entry at all on November 11 — and no model card to check it against. Treat the capability claims as Alibaba’s own description, not as measured fact.
The architecture hints are real but thin. Qwen3.8-Flash-Next, open-sourced on August 28, 2026, is described by Qwen’s own materials as an experimental preview of ideas expected in Qwen4, and community analyses describe it as a mixture-of-experts model with 176 billion total parameters and 6 billion active. Whether Qwen4-Max shares that design is unconfirmed — Alibaba has published no parameter count for Qwen4-Max itself.
How to try it today: the two doors
There are two ways in, and they are not the same door.
Door one: Qwen Chat (the free way). Qwen4-Max Preview now appears in Qwen Chat, Alibaba’s consumer chat surface. There is no per-token meter, no model ID to paste into a request, and no bill at the end of the month. You open the chat, pick the model, and talk to it. This is the fastest way to get a feel for the model — ask it to plan a project, review a document, or generate something visual — and for most readers it is the only door that makes sense.
Door two: the Token Plan (the developer way). The API route runs through Alibaba Cloud’s Token Plan, a credits-billed subscription for coding tools. This is the only API access that exists: Alibaba’s documentation says Qwen4-Max-Preview “is available through Token Plan only. Pay-as-you-go access is not supported.” There are three walls around it that catch people out:
- You need a Token Plan subscription. Individual plans are listed at $6 a month (Lite), $10 (Essential), $18 (Standard), and $43 (Pro). The plan bills in Credits from a monthly quota, not per token.
- You need a separate key and base URL. The Token Plan uses a dedicated base URL (
token-plan.ap-southeast-1.maas.aliyuncs.com), and an existing Model Studio API key does not open it. The November 4 quick-start guide describes OpenAI-compatible and Anthropic-compatible protocol options, with clients including Cursor, Codex, Claude Code, Qwen Code, and Qoder. - You cannot build it into a product. The Personal Edition terms limit Token Plan use to interactive use inside compatible coding and agent tools. Building Qwen4-Max into automation scripts or an app backend is forbidden — which, combined with the November 25 retirement, makes it useless as a production dependency in any case.
Early access also extends to QwenWork, Qoder, the Qwen app, and QwenStudio. If you already use one of those tools, the model may simply appear in the picker — Qwen Studio at chat.qwen.ai now lists Qwen4-Max-Preview next to Qwen3.8-Max.
The try-it-before-November-25 checklist
If you want to evaluate Qwen4-Max properly before it disappears, here is the practical sequence:
- Start in Qwen Chat today. It costs nothing and takes two minutes. Give it a task you already know the answer to — a code review, a document summary, a plan for a real project — so you can judge quality instead of vibes.
- Test the long context on something real. The 1M-token window is the model’s biggest claim. Feed it a large codebase or a long document set and ask a question that requires connecting facts from the beginning and the end. A model that holds a million tokens but loses the thread at token 200,000 is less useful than one with a smaller, sharper window.
- Test tool calls over a long task. Alibaba’s claim is sustained performance on long tasks with repeated tool calls. Run a multi-step coding task in Qwen Code or Qoder and watch whether quality holds up in step ten as well as step one.
- Only then consider the Token Plan. If the free chat tests impress you, the $6 Lite tier is the cheapest way to run the API version. Remember the promotion: Alibaba is cutting Qwen4-Max preview credit consumption to as little as 20% of the standard rate at launch — five times the usage for the same credits — with a further overnight discount bringing consumption to 8% of standard between 22:00 and 08:00 the following day (UTC+8).
- Write down your results. The model retires on November 25. Any notes you take now are the only record of how this version performed.
What it costs — and what it doesn’t
There is no per-token price for Qwen4-Max Preview, because there is no pay-as-you-go access. The only way to pay is the Token Plan subscription, and the only costs are the monthly tiers: $6, $10, $18, or $43.
That makes cost comparison with other models awkward by design. Qwen3.8-Max, the model Qwen4-Max is meant to replace, has a straightforward published price: $2.00 per million input tokens and $6.00 per million output tokens, and it has been generally available since August 3, 2026. Claude, GPT-6, and Gemini all publish per-token prices too. Qwen4-Max Preview publishes a subscription price and a credit quota, and the launch promotion temporarily cuts credit consumption to a fifth of the standard rate.
The practical consequence is simple: the cheapest honest way to try Qwen4-Max is free (Qwen Chat), and the cheapest API way is the $6 Lite tier with the launch promotion running. Do not try to calculate a per-token cost for this preview — the numbers do not exist yet, and any figure you see quoted today is someone’s estimate, not Alibaba’s price.
One more cost warning matters. Token Plan credits are a subscription quota, not a bank balance you keep forever. If you buy a plan to test Qwen4-Max and the model retires on November 25, your remaining credits still work for the other models on the plan — Qwen3.8-Max, Qwen3.8-Flash, and others are listed alongside it — but the thing you bought the plan to test will be gone.
What nobody knows yet — the honest box
This is the part of the launch that most coverage skips, and it is the part that matters most. Here is everything Alibaba has not published about Qwen4-Max:
- No benchmarks. No score on any public benchmark or leaderboard. No comparison against Claude, GPT-6, Gemini, or its own predecessor.
- No model card. No published documentation of training data, safety evaluations, or known limitations.
- No parameter count. Alibaba has not said how big the model is.
- No per-token price. There is no pay-as-you-go rate to compare against other models.
- No weights. Nothing to download or self-host.
- No date for the full release. No date for Qwen4-Plus, Qwen4-Flash, or the open-weight 27B tier either.
The early third-party evidence is thin and mixed. One early test from AIHubMix on November 11 reported a 300-token reply that came back empty — one test, from one provider, on day one, and not a verdict on the model. Community reaction in the first hours was skeptical in the same way: on a community forum thread tracking the launch, one early tester wrote that what they had seen so far was “actually pretty bad” and that the missing benchmarks were “weird,” while others compared it unfavorably with Claude’s Opus tier. Those are first-impression reactions from anonymous testers, not measurements — but they capture the mood honestly: a flagship launch with no numbers invites doubt, and Alibaba could have prevented that doubt by publishing even one benchmark.
The fair way to hold this in your head is: Qwen4-Max Preview is a real model you can use, and every claim about how good it is remains unproven until someone independent measures it — or until Alibaba publishes numbers of its own.
Qwen4-Max vs Qwen3.8-Max: the model it replaces
The most useful comparison available today is not against Claude or GPT-6. It is against Qwen3.8-Max, the current flagship that Qwen4-Max is meant to succeed. That comparison is possible because Qwen3.8-Max is a finished, documented, generally available product.
| Qwen3.8-Max | Qwen4-Max Preview | |
|---|---|---|
| Status | Generally available since August 3, 2026 | Preview only, Nov 11–25, 2026, then retired |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Inputs | Text, image, video | Text, image, video |
| Price | $2.00 / $6.00 per 1M input/output tokens | Token Plan subscription only, no per-token price |
| Benchmarks | Published (e.g. 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, 93.0 on PaperBench, 74.8 on CoWorkBench, per community reporting) | None published |
| Open weights | No | No |
| Safe to build on | Yes — a stable, priced API | No — retires November 25 |
Read that table carefully, because it tells you the real story of this launch. On the one spec both models publish — context window — they are identical. On everything else that matters for choosing a model, Qwen3.8-Max is the known quantity and Qwen4-Max is the unknown one. Alibaba says the new model beats the old one on world knowledge, code generation, and long tasks with repeated tool calls, but until there is a number attached to that claim, Qwen3.8-Max remains the model you can actually plan around.
That is not a criticism of Qwen4-Max. It is a preview. Previews are supposed to be unfinished. The mistake would be treating a preview like a product — and the November 25 retirement date is Alibaba’s own way of telling you not to.
Should you build anything on Qwen4-Max?
For most readers, the answer is no — not yet, and not on this version. Here is the decision framework, by who you are:
If you are a freelancer or small business owner curious about AI tools: try it in Qwen Chat for free. Use it for a real task — drafting, planning, reviewing — and judge it the way you would judge any new tool: did it save you time, and was the output good enough to use? If yes, keep it in mind for the full Qwen4 release. If no, you lost twenty minutes, not a budget.
If you are a developer evaluating models for a project: run your comparison now, inside the window, and record the results. Test the long context and the tool-call stamina, because those are the claims that would matter for your work. But do not write production code against qwen4-max-preview. The model ID stops working on November 25, the Token Plan terms forbid backend use, and there is no migration path announced to whatever replaces it.
If you run an agency or a product that calls AI models in production: stay on Qwen3.8-Max or your current provider. A two-week preview with no price and no benchmarks is not a vendor decision; it is a research project. The time to evaluate Qwen4 seriously is when Alibaba publishes the full release with weights or a stable API, a price, and a benchmark — whichever comes first.
If you are watching the open-weights question: the 27B tier is the one to watch. A 27-billion-parameter open-weight model in the Qwen4 family would be directly usable by small teams with a single good GPU, and it is the tier the local-model community has been most excited about since Apsara. It has no date yet. When it lands — if it lands — that is the Qwen4 release that changes things for independent builders.
What to watch next
Three things will tell you whether Qwen4 becomes a real option or stays a two-week curiosity:
- Benchmarks before November 25. If Alibaba or an independent evaluator publishes scores for Qwen4-Max Preview before it retires, those numbers become the baseline for judging the full release. If the window closes with no numbers at all, treat the silence as information.
- The sibling models. Qwen4-Flash, Qwen4-Plus, and the 27B tier are still unconfirmed. The 27B matters most for small teams; Flash matters most for cost-sensitive API use. Any of them landing with weights would be a bigger story than this preview.
- The full release date. Alibaba has not given one. The preview’s retirement on November 25 does not imply a general-availability date — it only implies that this version ends. Watch for an announcement in late November or December, and do not assume the preview model and the eventual flagship are the same thing.
The November AI calendar is crowded — Anthropic’s new Usage Policy takes effect today, November 12, and OpenAI’s third-party gateway screening starts November 20 — so Qwen4-Max is competing for attention in a month full of AI deadlines. Our roundup of November’s AI price and policy deadlines tracks the dates that actually affect what you pay and what you can use. And if you are comparing model prices in general, the Gemini 4 Argon pricing change coming November 30 is the other November deadline worth knowing about, since its introductory pricing doubles at the end of the month.
Frequently Asked Questions
When did Qwen4-Max Preview launch?
November 11, 2026. The official Qwen account on X posted “Qwen4 is coming.” at 02:01 UTC (10:01 Beijing time), and Alibaba listed the model on its Token Plan subscription the same morning, with early access in QwenWork, Qoder, the Qwen app, and QwenStudio.
When does the Qwen4-Max Preview end?
November 25, 2026, at 24:00 (UTC+8). Alibaba’s documentation says the preview runs from 10:00 on November 11 to 24:00 on November 25, no changes will be made during that period, and the model will be retired after the preview ends.
How can I try Qwen4-Max for free?
Open Qwen Chat and select Qwen4-Max-Preview from the model picker — no subscription, API key, or payment needed. That is the fastest free route. The API route requires a Token Plan subscription, starting at $6 a month for the Lite tier.
How much does Qwen4-Max cost?
There is no per-token price, because pay-as-you-go access is not supported. The only API access is through Alibaba Cloud’s Token Plan, a credits-billed subscription with tiers at $6, $10, $18, and $43 a month. A launch promotion cuts Qwen4-Max preview credit consumption to as little as 20% of the standard rate, and to 8% overnight (22:00–08:00 UTC+8).
Is Qwen4-Max open source?
No. There are no published weights, no model card, and no download. The open-weight tier of the Qwen4 family is the 27B model named at Apsara in September, and it has no release date yet.
Is Qwen4-Max better than Qwen3.8-Max?
Alibaba says it beats Qwen3.8 on world knowledge, code generation, and long tasks with repeated tool calls — but no benchmarks have been published, so the claim is unverified. Qwen3.8-Max remains the safer choice for real work: it is generally available, priced at $2.00/$6.00 per million input/output tokens, and it is not retiring on November 25.