Alibaba's Qwen4-Max Is Out for Two Weeks, Then It Is Retired. There Is No Benchmark, No Per-Token Price, and No Way to Use It Outside a Subscription.
Qwen posted that Qwen4 is coming on November 11 and opened qwen4-max-preview on Alibaba Cloud's Token Plan, with early access in QwenWork, Qoder, the Qwen app, and QwenStudio. Model Studio's documentation calls it the next-generation flagship multimodal model, text, image, and video in, text out, with a 1 million token context and up to 393,216 output tokens shared between thinking and the answer. The preview runs from November 11 to the end of November 25 Beijing time and is then retired. It is Token Plan only, with no pay-as-you-go, no per-token price, no parameter count, no benchmark table, and no weights. We read the Model Studio docs and the Qwen post.
By T.W. Ghost
Qwen posted "Qwen4 is coming" early today, November 11, 2026, and opened qwen4-max-preview on Alibaba Cloud's Token Plan, with early access in QwenWork, Qoder, the Qwen app, and QwenStudio. Alibaba's Model Studio documentation calls it the next-generation flagship multimodal model: text, images up to 16 megapixels, and video up to two hours or two gigabytes in, text out, with function calling, web search, structured outputs, and compatibility with the OpenAI Chat Completions and Responses APIs and the Anthropic Messages API. Context is 1 million tokens. Output is up to 393,216 tokens, shared between thinking and the final answer, with thinking on by default at the highest effort setting.
Everything a buyer would need to evaluate it is missing, by design. The preview window is 10:00 on November 11 to midnight at the end of November 25, Beijing time, which is 8 PM Central on November 10 to 10 AM Central on November 25; the docs say no changes will be made during that window, and the model is retired after it. Access is Token Plan only, a credit subscription, with pay-as-you-go not supported, so there is no per-token price to compute a cost per task from. There is no parameter count, no benchmark table, no weights, and no entry on Artificial Analysis. It is a public test with a hard stop. We read the Model Studio model page, the release notes, and the Qwen post. Here is what held up.
What Alibaba Published
| qwen4-max-preview (Model Studio, November 11, 2026) | |
|---|---|
| Window | 10:00 November 11 to 24:00 November 25, UTC+8; retired after; "no changes will be made during this period" |
| Access | Token Plan only; pay-as-you-go not supported; early access in QwenWork, Qoder, the Qwen app, QwenStudio |
| Context and output | 1M context; up to 393,216 output tokens shared by thinking and answer |
| Thinking | On by default; reasoning effort low, medium, or xhigh, default xhigh |
| Token accounting | max_tokens now equals max_completion_tokens and covers thinking, so budgets sized for answer-only output can truncate |
| Input | Text; images up to 16M pixels; video up to 2 hours or 2 GB |
| API compatibility | OpenAI Chat Completions and Responses; Anthropic Messages |
| Per-token price | None; credits on a subscription |
| Token Plan cost | Personal from about US$6 a month, new customers from US$3 (secondary, not checked on the pricing page) |
| Promo consumption rate | As low as 20% of standard, 8% overnight Beijing time (secondary, unverified) |
| Parameters, benchmarks, weights | None published |
| Artificial Analysis | No entry |
A two-week model is a demo, not a dependency. Anything built on this identifier stops working on November 25. The docs are explicit, which is to Alibaba's credit, and which makes integration work a waste.
Credits are not a price. A per-token figure is what lets a cost per task be computed and compared; a subscription with a consumption multiplier that changes at night does not. Nothing here can be placed on a cost board.
The default burns the fastest. Thinking on, at the highest effort, drawing from the same 393,216-token budget as the answer. A user who does not change the settings gets the most expensive configuration and may get a truncated answer if their output budget assumed answer-only accounting.
The name is the only thing with weight. "Qwen4" is a signal that the family is coming. The preview is evidence that something exists; it is not evidence of how good it is, and anecdotes on X are not benchmarks. Whether Qoder's stealth Qoduck is a Qwen4 model remains unconfirmed, even though both now sit in the same Qoder picker.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Model id; preview window; retirement; Token Plan only | Verified (Model Studio docs) |
| 1M context; 393,216 shared output cap; thinking defaults; token accounting change | Verified (Model Studio docs) |
| Input modalities and API compatibility | Verified (Model Studio text-generation page) |
| Early access surfaces | Verified (Qwen post) |
| Token Plan prices and promo consumption rates | Secondary; not checked on the pricing page |
| Parameter count, benchmark table, weights | None published (verified absence) |
| Artificial Analysis score | None |
| Qoduck is a Qwen4 model | Unconfirmed |
Who Should Use It
Try it on a Token Plan if you want an early look at the Qwen4 family and can throw the integration away on November 25.
Skip it for anything you ship, and skip any claim about how it compares to Kimi K3.1 Pro or anyone else until there is a number from someone other than a user on X.
The frontier is elsewhere: on the same morning Baidu shipped ERNIE 5.5 Flash Preview with a size, an API price, and a weights date, which is three more facts than Qwen4-Max has. Our model comparison has the current standings side by side.
Sources
- Alibaba Cloud Model Studio: Qwen4-Max Preview
- Alibaba Cloud Model Studio: newly released models
- Alibaba Cloud Model Studio: text generation models
- Qwen on X: Qwen4 is coming
- Sulat: Qwen4-Max Preview on Token Plan (secondary)
- bytefwd: Qwen4-Max Preview access and context guide (secondary)
*Not sure whether a two-week preview tells you anything about the model that follows it? Take the free 2-minute quiz and get matched. Then read about the stealth model in the same picker that will not say its name: Qoder's Qoduck.*