Qwen4-Max Preview: 14 Days to Try Alibaba's New Flagship AI

Alibaba opened Qwen4-Max Preview on 11 November 2026. It disappears on 25 November, giving your business fourteen days to try the first public model of the Qwen 4 family. The model ID is qwen4-max-preview, and Alibaba's documentation is blunt: no changes during the preview, and retirement when it ends.

The facts are short. Qwen4-Max Preview takes text, image and video input, has a 1,000,000-token context window, and can produce up to 393,216 tokens of output, with that ceiling covering its thinking and its final answer together.

What Qwen4-Max Preview actually is

Qwen4-Max Preview is the first Qwen 4 model anyone outside Alibaba can call. The family was named at Alibaba's Apsara conference on 22 September 2026, and the preview is the first piece of it to arrive: one tier, for two weeks, behind a subscription. Thinking is on by default. Alibaba's own announcement put it plainly: "Qwen4 is coming." A Qwen 4 release-date guide confirms the shape: one model in preview, the rest unannounced.

What Alibaba has not published

There is no parameter count, no model card, no weights, no benchmark score and no per-token price. An independent what-we-know-so-far rundown of the Qwen 4 launch reaches the same conclusion: nothing published yet lets you rank this model against its rivals.

The 14-day window: the dates that matter

The preview opened at 10:00 on Tuesday 11 November 2026 and closes at 24:00 on Tuesday 25 November 2026, Beijing time. Alibaba says the model will not change during those fourteen days. There is no successor date, so plan as though the endpoint stops on the 25th.

Start on 12 November and you have thirteen days; start on the 20th and you have five.

What the specs mean for real business work

A 1,000,000-token context window

You can hand the model a contract pack, a year of support tickets or a codebase in one go. But a large window is not reliable recall: Alibaba has published no evidence on long-range recall. Test it: bury the key fact mid-input. Our RAG chatbot guide explains why retrieval still matters.

A 393,216-token output ceiling, shared with thinking

Qwen4-Max Preview can produce up to 393,216 tokens in one response, but that total covers its reasoning and its final answer together. Thinking is on by default, so part of the budget goes to reasoning before you see any answer.

Text, image and video input, with thinking on by default

Multimodal input means the model takes screenshots, product photos and video clips as well as words. Test it on unglamorous work: a photographed invoice, a dashboard screenshot, a product photo with a defect. Judge it on your actual inputs.

Access and cost: Token Plan only

Qwen4-Max Preview is available only through Alibaba's Token Plan, a monthly subscription for coding and agent tools. Pay-as-you-go is explicitly not supported, and an ordinary Model Studio key will not open this model. The API route uses a dedicated Singapore base URL.

Where you can reach it

The model is also reachable in Qwen Chat at chat.qwen.ai, in QwenWork, in Qoder, in the Qwen app and in Qwen Studio. For a quick first look, use Qwen Chat. For a real evaluation, use the API route. Our chatbot-versus-agent guide sets out the difference.

What the $6 a month figure does and does not mean

Alibaba's published Token Plan tiers start at $6 a month. That is the subscription price, not a per-token price for Qwen4-Max: Token Plan bills in credits from a monthly quota, and Alibaba has published no per-token price at all. Record what each test workload costs in credits. Our AI calling agent cost guide and WhatsApp chatbot cost guide make the same point: the subscription price is not the running cost.

A 14-day evaluation plan: what to test

Fourteen days is enough to evaluate a model properly if you plan the fortnight. The plan below covers the three things Alibaba says this model is for: long-context agent tasks, multimodal reasoning and coding. Run your own work through it and score each test pass or fail. Score each test on accuracy, recall, completeness, multimodal reliability, credit burn and failure modes. Our AI agents and automation development service does this kind of evaluation for clients.

Days 1 to 4: long-context agent tasks

Give the model a real multi-step job: triage a week of support emails against your help-centre documents, or reconcile a supplier statement against your invoices. Score three things: did it finish without you splitting the input, did it cite the right documents, and did it hold detail from the middle? Our AI agents for small business automation guide lists the jobs worth automating first.

Days 5 to 9: multimodal reasoning

Feed it a photographed receipt, a dashboard screenshot, a product photo with a defect, or a short screen recording of a bug. Test the messy inputs and record where it guesses instead of saying it cannot tell. Our n8n vs Make vs Zapier comparison shows how a model like this slots into an existing workflow.

Days 10 to 14: coding

Give it a real task from your backlog: a bug fix, a small feature, or a test suite for an untested module. Judge it like a junior developer's pull request: does it run, does it handle the edge cases, and did it change anything it should not have touched? Our Claude Fable 5.2 coverage and Gemini 4 Argon pricing guide show what coding models cost elsewhere.

Why you must not build production on it

Do not build production systems on Qwen4-Max Preview. The model retires on 25 November 2026, Alibaba has announced no successor date, and nothing promises your prompts or integrations will carry over. Anything wired into a live workflow will break within two weeks. Prototypes and evaluations are the right use. Production is not.

A preview with no published benchmarks cannot be judged on agent ability yet. Our GDPx agent benchmark post showed even the best models score only around 12% on long-horizon agent tasks. Our Meituan LongCat-2.5 coverage made the same point. Our AI calling agent runs on models with published track records, and our how to build an AI voice agent guide explains the production discipline a preview cannot offer.

Qwen4-Max Preview against the models you already use

Without benchmarks, a comparison table would be guesswork. Here is an honest one instead.

• Qwen3.8-Max, the model it succeeds: callable since 3 August 2026 and the stable, documented option. Qwen4-Max Preview is newer and unproven. If you need something that will still be there in December, Qwen3.8-Max is safer.
• Claude Fable 5.2: a flagship coding model with published pricing, while Qwen4-Max Preview has no per-token price at all. Test both on the same coding tasks and compare outputs side by side.
• Meta Muse Ember: a small, cheap model for high-volume routine agent work, while Qwen4-Max is positioned as a flagship. Our Meta Muse Ember coverage explains where cheap models change the maths.
• Gemini 4 Argon: a November flagship with published pricing and a deadline to build before it doubles. Qwen4-Max Preview has a retirement date and no price.

Qwen4-Max Preview is the only model on this list you cannot plan around: a published end date, no published price, no published score. That makes it a model to evaluate, not adopt. If you want a second pair of eyes on your results, talk to Coheser. Our SaaS MVP cost guide covers what the system around the model costs to build. More from Coheser on AI agents and automation.

Frequently asked questions

What is Qwen4-Max Preview?

The first public model of Alibaba's Qwen 4 family, model ID qwen4-max-preview. It opened on 11 November 2026 with a 1,000,000-token context window and multimodal input, available only through Token Plan and Qwen's apps.

When does the Qwen4-Max preview end?

At 24:00 on 25 November 2026, Beijing time. Alibaba says the preview model is retired after it ends, with no announced successor date.

How much does Qwen4-Max Preview cost?

There is no published per-token price. Token Plan tiers start at $6 a month, billed in credits from a monthly quota. That $6 is the subscription price, not a per-token price.

Can I use Qwen4-Max Preview in production?

No. The model retires on 25 November 2026 with no announced replacement, and Alibaba has published no benchmark scores. Evaluate it for fourteen days, then wait for the full release.