Qwen4-Max Preview Explained: Alibaba’s 14-Day Flagship Test With a 1M-Token Context
Alibaba has finally put a Qwen 4 model in developers’ hands — but only just, and only for two weeks. The Qwen4-Max preview went live on November 11, 2026, as the flagship tier of the Qwen 4 family, and it will be retired on November 25. There are no benchmark scores, no per-token price, and no downloadable weights.
What there is: a 1-million-token context window, up to 393,216 output tokens, text, image and video input, and a subscription-only route in through Alibaba’s Token Plan. If you build with AI models, this is a short, sharp evaluation window — and this guide explains exactly what the Qwen4-Max preview is, what it costs, how it compares with Qwen 3.8-Max and Claude Opus 5.5, and what to do before the model string stops working.
Last updated: November 15, 2026. All facts in this article are drawn from Alibaba’s own documentation and launch materials as reported by independent trackers; where a figure is unverified, we say so plainly.
Table of Contents
- What Is the Qwen4-Max Preview?
- Qwen4-Max Preview Specs at a Glance
- How Access to the Qwen4-Max Preview Works
- Qwen4-Max Preview Pricing: What It Costs in the US
- Qwen4-Max Preview vs Qwen 3.8-Max
- Qwen4-Max Preview vs Claude Opus 5.5
- What Alibaba Has Not Told Us About the Qwen4-Max Preview
- Three API Changes That Will Trip Up Your Code
- How to Try the Qwen4-Max Preview Before November 25
- Should You Build on the Qwen4-Max Preview?
- What Comes After the Preview: Qwen4 Plus, Flash and 27B
- Frequently Asked Questions
What Is the Qwen4-Max Preview?
The Qwen4-Max preview is the first usable model from Alibaba’s Qwen 4 family, and it is a preview in the strictest sense of the word. Alibaba announced the Qwen 4 family at its Apsara Conference in September 2026, naming four tiers — Max, Plus, Flash and a 27B model — and saying the new generation was built on a next-generation architecture. For seven weeks after that, Qwen 4 existed only as a stage slide and a tease. On November 6, Alibaba’s official Qwen account posted a riddle about a fish called Kun, and technical leadership confirmed a release was coming the following week. Then, on November 11, the top tier quietly became callable.
Alibaba describes the Qwen4-Max preview as its next flagship multimodal model, built for coding, office productivity, front-end design and complex agent tasks. The launch post makes the same pitch in plainer terms: deeper knowledge, stronger coding, and better performance in agent workflows and professional software. In other words, this is not a chatbot demo. It is aimed squarely at developers and at the agentic workflows that now dominate AI spending — long tasks with repeated tool calls, code generation, and sustained multi-step work.
But the shape of the release is unusual, and it matters more than the headline specs. The Qwen4-Max preview is available only through Alibaba’s Token Plan, a credits-based subscription for coding tools. There is no pay-as-you-go API option. There is no model card, no technical report, no leaderboard table, and no downloadable weights. Alibaba has published no benchmark scores at all, and as of November 12 no independent lab had published measured results either. The company that once open-sourced a 2.4-trillion-parameter flagship has released its most anticipated model as a subscription line item with a name and almost nothing else.
The window is equally strict. The Qwen4-Max preview runs from 10:00 on November 11 to 24:00 on November 25, 2026, Beijing time — that is 9:00 pm Eastern on November 10 to 11:00 am Eastern on November 25. Alibaba says it will make no changes to the model during that period, and that the preview model will be retired when the period ends. Anything you build against the model ID qwen4-max-preview needs a new model string by November 25. That is the single most important practical fact about this release: it is a two-week evaluation window, not a product you can plan a roadmap around.

Qwen4-Max Preview Specs at a Glance
Here is everything Alibaba has actually published about the Qwen4-Max preview, in one place. Anything missing from this table is missing because Alibaba has not published it — not because we left it out.
| Spec | Qwen4-Max Preview |
|---|---|
| Model ID | qwen4-max-preview (billing renders as qwen4.0-max-preview) |
| Status | Preview, November 11 to November 25, 2026 (Beijing time) |
| Context window | 1,000,000 tokens |
| Maximum output | 393,216 tokens, shared by thinking and the final answer |
| Inputs | Text, images and video |
| Thinking | On by default, with low, medium and xhigh effort levels (xhigh is the default) |
| Tool use | Function calling, web search and built-in tools |
| Structured output | JSON Object and JSON Schema |
| API compatibility | OpenAI Chat Completions and Responses, Anthropic Messages |
| Access | Token Plan only; no pay-as-you-go |
| Parameter count | Not published |
| Open weights | Not released |
| Benchmarks | None published |
| Per-token price | None published |
Two numbers in that table deserve a closer look, because they are the ones every summary of the launch leads with. The first is the 1-million-token context window. That matches Claude Opus 5.5, which also offers 1M tokens, and it puts the Qwen4-Max preview in the top tier of models for long-document work — whole codebases, long legal files, and research corpora fed in a single pass.
The second is the 393,216-token output ceiling. That figure is shared between the model’s reasoning and its final answer, which changes how you should budget for it. A long reasoning trace eats the same budget as the answer itself, so the thinking-effort setting becomes your first cost control. We explain that in detail in the API section below.
Video input is the other genuinely new capability worth noting. The Qwen4-Max preview accepts text, image and video input, which the previous Qwen flagship did not offer in the same form. For front-end design work and for agent tasks that involve watching a screen recording or a product video, that is a real functional difference — though, again, without published benchmarks there is no independent measure of how good the video understanding actually is.
How Access to the Qwen4-Max Preview Works
There are three routes into the Qwen4-Max preview during its two-week window, and they serve very different audiences. Understanding which one you are on matters, because the limits and the billing are different on each.
Token Plan is the developer route. Token Plan is Alibaba’s credit-based subscription on its QwenCloud platform. You subscribe, receive a dedicated API key and a base URL, and set the model ID to qwen4-max-preview in a supported tool. The plan page lists Qwen Code, Cline, Claude Code, Cursor, OpenCode, Codex, Kilo CLI and OpenClaw as supported tools. Note the base URL: the Token Plan API runs on a Singapore-region endpoint, token-plan.ap-southeast-1.maas.aliyuncs.com, which is a different base URL from the standard Alibaba Cloud international API. If you point an existing client at the usual endpoint with the new model string, it will not resolve.
Pay-as-you-go API access is explicitly not supported for the preview model — the model page states that directly.
Alibaba’s own apps are the no-key route. The launch post names early access in QwenWork, Qoder, the Qwen app and QwenStudio. These are the product doors: no API key required, no base URL to configure, and no credit accounting to manage. QwenWork is Alibaba’s agent workspace, Qoder is its coding environment, and QwenStudio is its model playground. For a US developer who simply wants to feel what the model does on a real task, this is the fastest way in.
Qwen Chat is the browser route. As of November 12, the model picker at Qwen Chat lists Qwen4-Max-Preview, described as a preview of the flagship Qwen4 model. Alibaba has not published usage limits for it there. This is the route for a quick test — paste a hard prompt, watch the reasoning, and compare the answer with the model you use today. It is not the route for production traffic, and it is not the route for automation: Alibaba describes the preview as interactive use only, with no app backends or automation.
One more access detail matters for US users. The Token Plan preview is Singapore-region only. That has two practical consequences. First, latency from the US East Coast will be higher than a US-hosted API, which matters for interactive coding-agent loops where every round trip adds up. Second, any team with data-residency requirements should check whether routing prompts through a Singapore endpoint is acceptable before running proprietary code through the preview. Alibaba has not published a US-region option for the Qwen4-Max preview.

Qwen4-Max Preview Pricing: What It Costs in the US
There is no per-token price for the Qwen4-Max preview, and that is not an oversight in this article — Alibaba simply has not published one. During the preview, the only paid route is the Token Plan subscription, where requests draw down monthly Credits instead of being billed by the token. That makes the Qwen4-Max preview unusually hard to cost for a production workload, and it is the main reason to treat it as an evaluation exercise rather than a migration target.
Here are the Token Plan tiers as shown on November 12, in US dollars. These are launch-event prices, and the page shows higher list prices alongside them, so expect them to rise after the launch event ends.
| Plan | Price per month | Credits per month | Concurrent agents |
|---|---|---|---|
| Lite | $3 | 20,000 | 1 to 2 |
| Essential | $5 | 45,000 | 2 to 3 |
| Standard | $18 | 80,000 | 3 to 4 |
| Pro | $43 | 200,000 | 6 to 8 |
The list prices shown next to those launch prices are $8, $16, $25 and $51 a month, and the page says the Lite and Essential discounts are for new users. In the plan’s model list, qwen4-max-preview carries a limited-time 80% off tag, so preview usage draws down Credits at a reduced rate.
In practice, that means the $3 Lite tier is a genuinely cheap way to run a serious evaluation: 20,000 Credits at an 80% discount goes a long way on test prompts. But Alibaba publishes no credits-per-token rate for the Qwen4-Max preview, so you cannot convert Credits into a per-token cost or compare it cleanly with a per-token API. The honest budgeting statement is this: you know the monthly subscription cost, and you do not know the per-token cost.
One third-party route does publish a per-token figure, and it is worth knowing about with the right caveats. AIHubMix, an API aggregator, says it is the first platform to offer the model as a per-token API, serving it since launch day as coding-qwen4-max-preview at $0.353 per million input tokens and $1.059 per million output tokens, and it labels the channel a preview service that “may be unstable and is recommended for testing and evaluation”. Those are AIHubMix’s prices, not Alibaba’s, and they reflect an aggregator reselling access rather than an official rate card.
Still, they give a rough sense of the order of magnitude: if Alibaba’s eventual per-token price lands anywhere near that range, Qwen4-Max would be dramatically cheaper than Claude Opus 5.5, which costs $4.00 per million input tokens and $20.00 per million output. That comparison is speculative, though — Alibaba has said nothing, and the aggregator price is not an official one.
For a reference point on the previous generation, Qwen 3.8-Max costs $2.00 per million input tokens and $6.00 per million output on Alibaba Cloud’s international region. Alibaba has not said whether Qwen 4 Max will land near those rates.
One budgeting detail does carry over to any future price, though: thinking counts against the output limit. On the Qwen4-Max preview the output ceiling of 393,216 tokens is shared between the model’s reasoning and its final answer, and thinking is on at the highest effort level by default. Long reasoning traces are billed as output on most APIs, so the effort setting is the first cost control to learn. If you run the preview at the default xhigh effort on long tasks, expect the reasoning trace to consume a large share of your output budget before the answer even begins.
Qwen4-Max Preview vs Qwen 3.8-Max
The most useful comparison for most readers is not against a rival lab’s flagship. It is against Qwen 3.8-Max, the Qwen flagship you can call on a regular API today. That is the model most teams will keep using while the Qwen4-Max preview runs, and it is the baseline any evaluation should measure against.
| Comparison | Qwen4-Max Preview | Qwen 3.8-Max |
|---|---|---|
| Status | Preview until November 25, 2026 | Generally available since August 3, 2026 |
| Context window | 1M tokens | 1M tokens |
| Maximum output | 393,216 tokens (thinking + answer) | Lower; check current docs |
| Parameters | Not published | 2.4T total, about 95B active |
| API access | Token Plan subscription only | Pay-as-you-go API |
| Price | Credits, no per-token rate | $2.00 / $6.00 per million tokens |
| Open weights | Not released | Published August 12–13, 2026 |
| Benchmarks | None published | Available from Alibaba and independent labs |
| Inputs | Text, image and video | Text and image |
Alibaba says the Qwen4-Max preview beats Qwen 3.8-Max on world knowledge, code generation and sustained long tasks with repeated tool calls. That is Alibaba’s own claim, and it is the only performance statement the company has made. It is also, notably, a claim about exactly the workloads where a preview evaluation is most valuable: not single-turn question answering, but long-horizon work where an agent calls tools repeatedly and has to keep track of what it has already done. If that claim holds, the Qwen4-Max preview would be most valuable to teams running coding agents and multi-step automation — the same audience the Token Plan is built for.
The practical read is simple, and it is the same one independent trackers have reached: the preview is for evaluation, and 3.8-Max is for production. Qwen 3.8-Max has published weights, published benchmarks, a per-token price, pay-as-you-go access, and no retirement date. The Qwen4-Max preview has none of those things, and it disappears on November 25. Run your comparison during the window, keep your production traffic on 3.8-Max, and do not migrate anything that has to keep working on November 26.
There is also a family-history point worth making for readers who follow the Qwen line. The Qwen family has built its reputation on open weights: Qwen 3.8-Max shipped with published weights in August 2026, and earlier Qwen models were among the most widely self-hosted open models in the world. The Qwen4-Max preview breaks that pattern. It is closed-weight, subscription-only, and time-limited. Whether that is a permanent shift in Alibaba’s strategy or a one-off preview arrangement is one of the biggest open questions about Qwen 4, and the answer will probably only arrive with the general release — or with the 27B tier, which Alibaba has described as the open-weights tier of the family.
If you want more background on the Qwen family before you evaluate the preview, our Qwen-Image-2.1-Turbo explainer covers the image-generation side of the same model family, including the $0.014-per-image pricing that made it one of the cheapest ways to generate images in 2026.
Qwen4-Max Preview vs Claude Opus 5.5
The comparison readers keep asking for is against Claude Opus 5.5, Anthropic’s flagship, released on September 22, 2026. It is a natural pairing: both models offer a 1-million-token context window, both are aimed at coding and long-horizon agent work, and both are the flagship of their respective families. But it is also a comparison that cannot be scored yet, and any page that ranks the two models on quality today is guessing. Here is what can be compared honestly.
| Comparison | Qwen4-Max Preview | Claude Opus 5.5 |
|---|---|---|
| Status | Preview, retires November 25, 2026 | Generally available; retirement not sooner than September 22, 2027 |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Maximum output | 393,216 tokens (thinking + answer) | 128,000 tokens (300,000 on the Batches API with a beta header) |
| Inputs | Text, images and video | Text and images |
| Price | Token Plan from $3/month; no per-token rate | $4.00 / $20.00 per million input/output tokens |
| Benchmarks | None published | Artificial Analysis Intelligence Index 59, ranked fifth of 74 models |
| Access | Token Plan (Singapore) + Qwen apps | Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry |
| Open weights | Not released | Closed |
On paper, the Qwen4-Max preview wins two categories outright: maximum output per response and input modalities. Its 393,216-token output ceiling is roughly three times Opus 5.5’s standard 128,000-token limit, and video input is something Opus 5.5 does not offer. For a task like generating a very long document, a large code refactor, or a multi-file front-end build in a single response, that output headroom is a genuine advantage — provided the quality holds up across all 393,000 tokens, which nobody has independently verified.
On availability, the comparison runs the other way, and it is not close. Claude Opus 5.5 is a generally available product with a published rate card, a retirement commitment of not sooner than September 22, 2027, and distribution through every major cloud. The Qwen4-Max preview is a fourteen-day experiment with no rate card and a fixed retirement date ten days from now. A team choosing a model for production work is not choosing between two flagships; it is choosing between a flagship and a demo.
The price comparison is where the most misleading claims are circulating, so it deserves care. The aggregator price of $0.353 per million input tokens and $1.059 per million output tokens, if it reflected Alibaba’s eventual rate, would make Qwen4-Max roughly 90% cheaper than Opus 5.5. But it does not reflect an official rate — Alibaba publishes no per-token price, and the aggregator is reselling preview access. The only honest price statement today is that the Qwen4-Max preview costs $3 a month to start evaluating, and its production price is unknown. Anyone quoting a per-token Qwen4-Max price as Alibaba’s price is repeating an aggregator’s number as if it were official.
For context on the wider flagship race, our GPT-6 Luna explainer covers OpenAI’s flagship from the same generation, and our DeepSeek V4.1 Flash explainer covers the efficiency-first challenger that competes on price rather than scale. Together, the three articles map the current flagship landscape: OpenAI on capability, DeepSeek on efficiency, and Qwen 4 on a two-week try-before-you-commit window.

What Alibaba Has Not Told Us About the Qwen4-Max Preview
The most honest section of any Qwen4-Max preview article is the list of what is missing, because the gaps are larger than usual for a flagship launch. Alibaba’s launch post and model documentation contain no benchmark numbers, and as of November 12 no independent lab had published measured results either. That matters because numbers are already circulating. Scores posted on X and Reddit do not trace to Alibaba or to a published leaderboard entry, so treat them as unverified. The same goes for two anonymous models people have linked to Qwen 4 — Kalmia on the LM Arena leaderboard and Qoduck inside Qoder. Alibaba has not confirmed either one as Qwen4-Max, and neither should be cited as its score.
The parameter count is unpublished. Qwen 3.8-Max is 2.4 trillion parameters with about 95 billion active per token, but any Qwen 4 Max size you see quoted is a guess until Alibaba publishes one. The model card does not exist. There is no licence, because there is nothing to download. The general release date is unannounced. Nothing new has been said about Qwen 4 Plus, Flash, or the 27B tier since the September stage reveal, and the four tier names themselves come from a captured stage slide — Alibaba’s published press release states only that Qwen 4 is in training.
There is also no independent measurement of speed. Artificial Analysis, the tracker that scores most frontier models, has no page for the Qwen4-Max preview. As of November 11 there was no measured speed, no cost-per-task figure, no index score and no provider count for it. The preview is too new, and at fourteen days it may never get an entry. BenchLM’s Qwen4-Max profile, updated on launch day, still marked the context window as coming soon and recorded zero benchmark coverage out of more than eight hundred tracked evaluations.
That does not mean there is no evidence at all. Hands-on tests are appearing, and they are the only performance data available. One independent tester paid for the Token Plan and gave the Qwen4-Max preview the same job he had given another flagship a month earlier — build an interactive 3D Noah’s Ark in a single HTML page, with animal pairs, live weather, sea mist, a captain’s log, procedural sound and a wireframe view. The Qwen4-Max preview produced the best result of the four models he tested, with the most complete feature set.
That is one hands-on test, not a benchmark, and the tester says so himself. But it is the kind of evidence a two-week preview is designed to produce: real tasks, run by real users, with the results published while the window is still open.
Another early report is more cautionary, and it is worth knowing before you spend Credits. One tester found that a 300-token output limit returned an empty reply — because on this model, the output limit covers thinking and the answer together, and the thinking alone consumed the whole budget. That is not a model failure; it is the API behaviour change described in the next section. But it shows how easily a first test can mislead if you carry over settings from an older model.
Three API Changes That Will Trip Up Your Code
Three behaviours are different from earlier Qwen models, and they will trip up existing code if you swap the model string and nothing else. All three are documented by Alibaba, and all three are easy to handle once you know about them.
First, the output limit now includes thinking. Starting with the Qwen4-Max preview, max_tokens means the same thing as max_completion_tokens: the combined budget for reasoning and the final answer. A limit that was sized for the answer alone can cut the response short — or, as the tester above discovered, return nothing at all. Earlier Qwen models keep their old behaviour, so this is a per-model difference, not a platform-wide change. If you are migrating a client from Qwen 3.8-Max, raise the output limit substantially, and size it for thinking plus answer, not for the answer alone.
Second, thinking is on by default at the highest level. The reasoning_effort setting takes low, medium or xhigh, and xhigh is the default. Alibaba’s documentation says not to set reasoning_effort and thinking_budget together — pick one. For most evaluation work, the practical move is to start at medium effort on routine tasks and reserve xhigh for the hard ones, because xhigh thinking consumes the shared output budget fastest. If your test prompts are short and your answers are coming back truncated, the effort setting is the first thing to check.
Third, multi-turn calls keep the reasoning. In Chat Completions, earlier reasoning is preserved across turns by default. The earlier reasoning goes back in the reasoning_content field of the assistant messages, not merged into the content. That is good for context — the model remembers how it got to its previous answer — but it means multi-turn conversations grow in token size faster than they did on older models, because the reasoning trace is carried forward each turn. For long agent sessions with dozens of turns, that growth is worth watching.
The rest is familiar. The Qwen4-Max preview works with the OpenAI Chat Completions and Responses formats and the Anthropic Messages format, so an existing client connects with a base URL, a key and the model string. The one exception is the base URL itself: Token Plan uses the Singapore endpoint, not the standard international endpoint, so update the base URL as well as the model string.

How to Try the Qwen4-Max Preview Before November 25
The window is short and the model string expires, so use the preview to answer one question: is Qwen 4 Max better on your workload than the model you run today? Here is a practical plan that fits inside the remaining days.
Step 1: pick your route. If you want a feel for the model, open Qwen Chat and select Qwen4-Max-Preview from the model picker. No subscription, no key, no configuration. If you want to test it inside your own coding tools, subscribe to Token Plan — Lite at $3 a month is enough for an evaluation — and set the model ID to qwen4-max-preview with the Singapore base URL in Qwen Code, Cline, Claude Code, Cursor, OpenCode, Codex, Kilo CLI or OpenClaw. If you want the agent-workspace experience, try the early access in QwenWork, Qoder, the Qwen app or QwenStudio.
Step 2: set up the comparison before you spend Credits. Pull a fixed set of 20 to 50 prompts from your own product — real tasks, not benchmark questions — run them on your current model, and save the outputs and token counts as a baseline. If Qwen 3.8-Max is that baseline, it is available on a regular pay-as-you-go API, so the baseline run and the competitor run are a model-string change in one client. Then run the same fixed set through the Qwen4-Max preview and compare outputs and token counts side by side. That tells you more about your workload than a leaderboard will, and it is the only evidence available until Alibaba or an independent lab publishes results.
Step 3: mind the output budget. Set max_tokens generously — remember it covers thinking plus answer — and start reasoning effort at medium rather than the xhigh default. If a response comes back empty or cut short, the budget is the first suspect, not the model. Keep the model ID in configuration, not in code: the preview ID goes away on November 25, and whatever replaces it will be a different string.
Step 4: test the things only this model offers. Try a video-input task, since that is the capability the previous flagship lacks. Try a very long output — a multi-file front-end build or a long document — since the 393,216-token ceiling is the headline advantage. And try a sustained agent task with repeated tool calls, since that is the workload Alibaba claims the Qwen4-Max preview is best at. A single-turn question will not tell you much; a long task will.
Step 5: note the calendar. The preview ends at 24:00 Beijing time on November 25, 2026 — 11:00 am Eastern that day, and 4:00 pm UTC. Alibaba says the preview model is retired after that. Plan your evaluation to finish a day early, so a last-day rush does not collide with the cutoff.
Should You Build on the Qwen4-Max Preview?
The short answer is: evaluate on it, do not build on it. The Qwen4-Max preview is a two-week test drive of a flagship that may or may not become a product you can buy. That distinction should drive every decision you make about it.
Try it if you run coding agents or long-horizon agent workflows and want to know whether Alibaba’s next flagship is better on your tasks than your current model. Try it if you need video input or very long outputs and want to test those capabilities before committing. Try it if you are a Qwen 3.8-Max user and want to know what the next generation looks like. At $3 for the Lite Token Plan tier, the evaluation is cheap, and the 80% preview discount on Credits makes it cheaper still.
Do not build on it if your product has to keep working on November 26, because the model string stops working that day. Do not migrate production traffic to it, because there is no per-token price, no service commitment, and no retirement date beyond the preview itself. Do not sign a customer commitment that depends on its behaviour, because Alibaba says the model will not change during the preview — but it has said nothing about what the general-release model will do, or when it will arrive. And do not quote its performance to a client or a manager, because there are no published benchmarks and no independent measurements to stand behind.
The most useful way to think about the Qwen4-Max preview is as an option, not a product. Alibaba is giving developers two weeks to test a flagship before it decides — or announces — what the general release looks like. The developers who use those two weeks well will have real data on their own workloads when the general release arrives. The developers who treat the preview as a launch will have production code pointing at a dead model string on November 26.
What Comes After the Preview: Qwen4 Plus, Flash and 27B
The Qwen4-Max preview is one tier of four, and the other three are still unreleased. At the Apsara Conference in September, Alibaba named Max as the flagship, Plus as the medium tier, Flash as the lightweight tier, and a 27B model as the on-device tier. Reporting in November described the four tiers as releasing sequentially across November and December. As of November 15, only Max has appeared, and only as a preview.
The 27B tier is the one to watch for a different reason. Alibaba has described it as the open-weights tier of the family, and the Qwen family’s reputation was built on open weights. If the 27B ships with weights under an open licence, it would restore the pattern that made Qwen 3.8-Max and its predecessors so widely self-hosted — and it would give developers something the Max preview deliberately does not: a model they can run themselves, keep, and build on without a subscription or a retirement date. No release date has been announced for it.
The general release of Qwen 4 Max itself also has no announced date. The preview ends November 25, and there is a gap between that date and any general release that Alibaba has not addressed. It is possible the general release follows quickly, with a per-token price and a model card. It is also possible there is a pause while Alibaba digests preview feedback. Either way, the sensible move is to treat November 25 as a hard stop, keep your evaluation data, and watch for the general-release announcement rather than assuming continuity.
One more thing to watch: the “Kun” tease. Alibaba’s November 6 post — “In the Northern Ocean there is a fish, and its name is Kun” — was the first public signal that Qwen 4 was coming, and it set off a week of speculation about a model codenamed Kun. The Qwen4-Max preview is the answer to that tease, at least in part. Whether “Kun” refers to the whole Qwen 4 family, to the Max tier specifically, or to something else entirely is still unclear, and Alibaba has not said. It is a small mystery, but it is a reminder of how little has actually been disclosed about this generation so far.

Frequently Asked Questions
What is Qwen 4 Max?
The flagship tier of Alibaba’s Qwen 4 family. It is available now as the Qwen4-Max preview, a multimodal reasoning model with a 1M-token context that Alibaba describes as built for coding, office work, front-end design and agent tasks.
Is Qwen 4 Max out?
In preview, yes. The Qwen4-Max preview has been available since November 11, 2026. The general release has no announced date, and the preview model is retired on November 25, 2026.
How do I access the Qwen4-Max preview?
Subscribe to Alibaba’s Token Plan and set the model ID to qwen4-max-preview in a supported coding tool, or use the early access in QwenWork, Qoder, the Qwen app or QwenStudio. It is also listed in the Qwen Chat model picker.
How much does Qwen 4 Max cost?
There is no per-token price. The Token Plan that includes the preview starts at $3 a month for new users during the launch event, with Standard at $18 and Pro at $43, and usage draws down monthly Credits. An aggregator lists $0.353 per million input tokens and $1.059 per million output tokens, but that is the aggregator’s price, not Alibaba’s.
Are there Qwen 4 Max benchmarks?
Not yet. Alibaba published none with the preview, and no independent lab results had appeared as of November 12. Scores circulating on social media are unverified, and anonymous leaderboard models have not been confirmed as Qwen4-Max by Alibaba.
What is the Qwen 4 Max context window?
1M tokens, with up to 393,216 output tokens shared between thinking and the final answer.
How big is Qwen 4 Max?
Alibaba has not published a parameter count. Qwen 3.8-Max is 2.4 trillion parameters with about 95B active, but any Qwen 4 Max size you see quoted is a guess until Alibaba publishes one.
Does Qwen 4 Max have open weights?
No. Nothing has been released for the preview, and Alibaba has not said whether the Max tier will get open weights. The 27B is the tier Alibaba has described as open weights.
When does the Qwen4-Max preview end?
At 24:00 Beijing time on November 25, 2026, which is 11:00 am Eastern that day. Alibaba says the preview model is retired after that.
When do Qwen 4 Plus, Flash and 27B come out?
No dates. Alibaba named all four tiers in September and has only released Max, in preview.
Sources: Alibaba’s Qwen4-Max-Preview model documentation and launch materials, as reported in the Yotta Labs Qwen 4 Max specs guide and the OrcaRouter Qwen4-Max Preview vs Claude Opus 5.5 comparison; hands-on testing notes from the November 15 Qwen4-Max preview test. Benchmark and pricing figures attributed to Alibaba are Alibaba’s own; aggregator prices are the aggregator’s, not Alibaba’s.
Bottom line: the Qwen4-Max preview is real and usable, with a 1M-token context, video input and reasoning on by default — but it is a two-week preview behind a subscription, with no benchmarks, no per-token price and no weights. Treat it as an evaluation window. Run your own prompts through it before November 25, compare against the model you call today, and keep the model string in config.
